A post-processing method for false alarm filtering of behavior wear in a smart factory

Abstract

The present application belongs to the technical field of video monitoring identification, and provides a post-processing method for false alarm filtering of behavior and wearing in a smart factory, which comprises the following steps: constructing an image data set of employee behavior and wearing; constructing a detection network model for employee behavior and wearing behavior, and inputting the constructed image data set into the detection network model for training; intelligently detecting the behavior and wearing state of employees in the factory area by using the trained detection network model; analyzing the detection result of the detection network model for first filtering; and performing secondary filtering when one "human body" target frame corresponds to multiple "head" target frames. The present application solves the problem that the prior art cannot well avoid false alarms when an employee does not correctly wear a safety helmet under the conditions of light and shadow and personnel gathering, thereby reducing the accuracy of safety production alarm.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of video monitoring recognition, and more particularly to a post-processing method for intelligent factory behavior and wearing false alarm filtering. BACKGROUND

[0002] In the technical field of the chemical industry, the identification of non-standard behavior of employee behavior and wearing behavior has reached a high level of identification accuracy, which can accurately identify the non-compliant behavior of personnel wearing in the factory area, such as employees not wearing safety helmets, not wearing work clothes, and not wearing anti-throw shoes. Among them, the employee not wearing a safety helmet is the top priority for safe operation. It is worth noting that in some on-site scenarios, human shadows caused by sunlight and personnel gathering often cause shadow false alarms and repeated alarms. Therefore, in the process of promoting the intelligent construction and application of the emergency safety production industry, it is necessary to optimize the processing of this typical false alarm to avoid unreasonable operation and alarm.

[0003] The research route in the art is to collect a large amount of image data in the factory area for labeling, use a target detection model to extract the general features of each target through deep learning of the labeled image data, and improve the identification performance of the model through accumulation of data volume, so as to reduce the false alarm rate of personnel wearing. However, this method requires collecting specific false alarm images and requires a large amount of manpower to label the images.

[0004] Chinese patent document CN113610037A discloses a method for detecting occluded pedestrians based on head and visible region clues, which improves the performance of occluded pedestrian detection in complex road scenes by using head, visible region, and full-body information. A sharpening feature response region recommendation network (SRPN) is proposed to simultaneously detect the full body and head of RoIs. For the full body RoIs, a visible region attention guidance (VGA) module is added before the original full body detection branch Fast RCNN to enhance the visible region features and suppress the occluded region noise. A parallel head cascade detection branch is added to detect severely occluded pedestrians based on the head RoIs.

[0005] In summary, the prior art cannot well avoid false alarms when employees do not wear safety helmets under the conditions of light and shadow and personnel gathering, which reduces the accuracy of safety production alarms and affects production and work. How to find an efficient, time-saving, labor-saving, reliable, and stable method for handling non-compliant personnel wearing in an intelligent factory to improve the accuracy and reliability of operation and alarm is an urgent consideration for professionals in this field. SUMMARY

[0006] The present application aims to overcome at least one of the above-mentioned defects of the prior art, and provides a post-processing method for false alarm filtering of behavior and wearing in a smart factory.

[0007] The detailed technical solutions of the present application are as follows:

[0008] A post-processing method for false alarm filtering of behavior and wearing in a smart factory comprises the following steps:

[0009] S1, acquire historical monitoring images of a factory area, and construct an image dataset of employee behavior and wearing;

[0010] S2, construct a cascade-rcnn detection network model for employee behavior and wearing behavior, and input the constructed image dataset into the cascade-rcnn detection network model for training.

[0011] S3, use the trained cascade-rcnn detection network model to intelligently detect the behavior and wearing state of employees in the factory area, and the detection result includes a "human body" target frame and its target information, a "head class" target frame and its target information, and the "head class" target frame includes a "head" target frame and a "wearing safety helmet" target frame.

[0012] The target information includes target frame coordinate values and category confidence, and the target frame coordinate values include left upper corner point coordinate values and right lower corner point coordinate values. Filtering through the left upper corner point and the right lower corner point of the coordinate frame does not require large-scale modification and improvement of the model structure, does not require retraining and replacement of the model, has low actual application cost, and has wide applicability. The category confidence is the confidence of the target frame being the current detected category.

[0013] S4, first filtering:

[0014] An intersection ratio between the "human body" target frame and the "head class" target frame is obtained by analyzing the detection result of the cascade-rcnn detection network model, a filtering threshold is set, and first filtering is performed.

[0015] If the intersection ratio is greater than the filtering threshold, the "head class" target frame corresponding to the "human body" target frame is retained, i.e., it is considered that the "head class" belongs to the current "human body"; if the intersection ratio is less than or equal to the filtering threshold, the "head class" target frame corresponding to the "human body" target frame is filtered out.

[0016] If one "human body" target frame corresponds to one "head class" target frame, and the "head class" target frame is a "wearing safety helmet" target frame, the verification is ended, i.e., the "human body" is wearing a safety helmet; if the "head class" target frame is a "head" target frame, an alarm for non-compliance of wearing is pushed, i.e., the "human body" is not wearing a safety helmet.

[0017] S5, second filtering:

[0018] When one "human body" target frame corresponds to multiple "head" target frames, that is, the "human body" may have multiple heads, a second filtering is performed:

[0019] All "human body" target frames are traversed to calculate the left line distance of the "human body" target frame to the multiple "head" target frames corresponding to the "human body" target frame, and the middle line distance of the "human body" target frame to the multiple "head" target frames corresponding to the "human body" target frame, and the mean value of the left line distance and the middle line distance are calculated and composed into a distance list.

[0020] The left line distance is the straight line distance from the top left corner point of the "human body" target frame to the top left corner point of the "head" target frame.

[0021] The middle line distance is the straight line distance from the middle point of the top of the "human body" target frame to the middle point of the top of the "head" target frame.

[0022] S6, according to the distance list and the category confidence, it is judged whether to push the alarm of wearing non-compliance:

[0023] The category confidence of all "head" target frames corresponding to the "human body" target frame is obtained and arranged in size order.

[0024] Case one: if the "head" target frame closest to the "human body" target frame and the "head" target frame with the highest category confidence are the same, then the "head" target frame is retained.

[0025] Case two: if the "head" target frame closest to the "human body" target frame and the "head" target frame with the highest category confidence are not the same, then the difference in category confidence between the "head" target frame closest to the "human body" target frame and other "head" target frames is calculated and the absolute value is taken, and a difference threshold is preset.

[0026] If the absolute value of the difference in category confidence is less than the difference threshold 0.1, then the "head" target frame closest to the "human body" target frame is retained; otherwise, the "head" target frame with the highest category confidence is retained.

[0027] If the retained "head" target frame exists, the "head" target frame is pushed to the monitoring system and the alarm of wearing non-compliance is performed.

[0028] The S1 specifically includes: extracting frame images of historical videos of factory scenes, manually labeling the categories of each image, the categories including "human body", "head" and wearing requirements of intelligent factory, and using the labeled images as an image dataset.

[0029] The smart factory wear requirements include: "wearing a safety helmet", "wearing a tool", "wearing a coat", "wearing ordinary shoes", "wearing anti-hitting shoes", and the labeled image is used as an image data set.

[0030] The S4 analyzes the detection results of the cascade-rcnn detection network model to obtain the intersection ratio between the "human body" target frame and the "head class" target frame, specifically:

[0031] All "heads" and "safety helmets" target frames in the current "human body" target frame are obtained, and the next step of filtering is performed;

[0032] The intersection ratio is calculated by the following formula:

[0033] The intersection part of the "human body" target frame and the "head class" target frame is calculated as the left upper corner point A (x min ,y min ) and the right lower corner point B (x max ,y max ) coordinates: x min =max(box1[0],box2[0]);y min =max(box1[1],box2[1]);x max =min(box1[2],box2[2]);y max =min(box1[3],box2[3]);

[0034] The box1 refers to the coordinate list of the "human body" target frame, including; [x1 min ,y1 min ,x1 max ,y1 max ], wherein box1[i] is the i-th element in box1, i=0, 1, 2, 3; the box2 refers to the coordinate list of the "head class" target frame, including; [x2 min ,y2 min ,x2 max ,y2 max ], wherein box2[i] is the i-th element in box2, i=0, 1, 2, 3;

[0035] The left upper corner point A (x min ,y min ) and the right lower corner point B (x max ,y max ) of the intersection part of the frame can be determined by the size of the left upper corner point x and the right lower corner point y of the "human body" target frame and the "head class" target frame;

[0036] The area of the frame of the intersection part is calculated as:

[0037] intersection = max(0, x max -x min )*max(0, y max -y min ) (1);

[0038] The intersection ratio, i.e. the ratio of the frame of the intersection part to the frame of the "human body" target, is calculated:

[0039] box1_area = (box1[2] - box1[0] + 1) * (box1[3] - box1[1] + 1) (2);

[0040] cover_percentage = float(intersection_area) / float(box1_area) (3);

[0041] wherein box1_area refers to the area of the frame of the "human body" target, cover_percentage refers to the intersection ratio, and float refers to a decimal.

[0042] The calculation method of S5 is specifically as follows:

[0043] The obtained coordinate values of the target frame include the coordinate values of the top-left corner point of the target frame and the bottom-right corner point of the target frame;

[0044] The left line distance distance ， between the top-left corner point (x1 min , y1 min ) of the frame of the "human body" target and the top-left corner point (x2 min , y2 min ) of the frame of the "head class" target is calculated, and the middle line distance distance ，， between the middle point of the top of the frame of the "human body" target and the middle point of the top of the frame of the "head class" target is calculated.

[0045] The average distance distance of the two distances is calculated.

[0046] ;

[0047] ;

[0048] (4);

[0049] (5);

[0050] (6);

[0051] (7);

[0052] (8);

[0053] x1 ul is a horizontal coordinate value of the top middle point of the first target frame, x2 ul is a horizontal coordinate value of the top middle point of the second target frame, x1 um is a horizontal coordinate value of the top middle point of the first target frame, x2 um is a horizontal coordinate value of the top middle point of the second target frame.

[0054] The first target frame is a "human body" target frame, and the second target frame is a "head type" target frame.

[0055] Compared with the prior art, the present application has the following beneficial effects:

[0056] The post-processing method for intelligent factory behavior wearing false alarm filtering provided by the present application can accurately filter out false alarms of employees not wearing safety helmets due to light and shadow and personnel gathering by differentiating false and missed reports into two kinds and setting two filtering operations, avoids unreasonable operation and maintenance alarms, improves the accuracy of safety production alarms, and based on filtering processing of the recognition model result, does not need to modify and improve the model structure on a large scale, does not need to retrain and replace the model, and has low actual application cost. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a method flowchart of the present application.

[0058] Figure 2 is a simplified schematic diagram of shadow false alarm in embodiment 1 of the present application.

[0059] Figure 3 is a schematic diagram of shadow false alarm in embodiment 1 of the present application.

[0060] Figure 4 is a simplified schematic diagram of side-by-side walking false alarm 1 in embodiment 1 of the present application.

[0061] Figure 5 is a simplified schematic diagram of side-by-side walking false alarm 2 in embodiment 1 of the present application.

[0062] Figure 6 is a schematic diagram of side-by-side walking false alarm in embodiment 1 of the present application. DETAILED DESCRIPTION

[0063] The present disclosure will be further described below in conjunction with the drawings and embodiments.

[0064] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the present disclosure. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.

[0065] The embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0066] Embodiment 1

[0067] This embodiment takes the real-time video monitoring of a chemical plant area as an example. According to a certain time interval, a certain frame of video stream image is extracted from the monitored video stream and sent to the cascade rcnn target detection model used in the present application. The model outputs various behavior and wearing information of employees in the plant area at a certain time. The categories include “person”, “head”, “head_helmet”, “uniform”, “coat”, “shoe”, “protective_shoe”, etc. Among them, “person”, “head_helmet”, “uniform”, and “protective_shoe” categories are positive samples, which meet the safety production requirements and should not generate alarms. The remaining categories are negative samples, which need to generate alarms for non-compliant wearing.

[0068] According to the detection result of the behavior and wearing state of the present application, the false alarm situation of the safety helmet not being correctly worn is solved by post-processing and business logic judgment method, and the behavior and wearing state of the employee is obtained. The image of the non-compliant wearing behavior is pushed and alarmed. The detailed content is:

[0069] Original business logic strategy

[0070] For the target frame category detected by the model, the coordinate information and category confidence of “person” and “head” are extracted.

[0071] All “person” target frames are traversed, and the intersection ratio between the current “person” target frame and each “head” target frame is calculated. If the ratio is greater than the set threshold value 0.8, the “head” target frame is retained.

[0072] All retained “head” target frames are pushed to the platform for non-compliant wearing alarm.

[0073] In field application, the “head” target frame pushed by the original business logic strategy will produce the following two typical false alarm situations:

[0074] Case 1: Shadow false alarm. When the sun shines on the personnel, the model may mistakenly identify the head position in the shadow as a real “head”, resulting in false alarm. Case 1 shadow false alarm is as follows:Figure 2 , 3 As shown in FIG. 6, conf is the confidence, wherein, Figure 3 The left part is the prediction result of the model, Figure 3 The right part is the actual alarm graph, Figure 3 The middle figure is the correct wearing behavior, but the shadowed head is considered as not wearing a safety helmet, resulting in false alarm;

[0075] Case two: false alarm or missed alarm when people walk side by side. When people walk in a row, the image presents two-dimensional information, resulting in misplacement of the human body, and a "human body" target frame contains the "head" of two people, resulting in repeated false alarm or individual missed alarm. The schematic diagram of false alarm or missed alarm when people walk side by side mainly includes the following two kinds, as shown in FIG. 7 and FIG. 8, wherein conf is the confidence, wherein, Figure 4 , 5 , Figure 6 The left part is the prediction result of the model, Figure 6 The right part is the actual alarm graph, Figure 6 The middle two figures both have incorrect wearing behavior, but one of the two people blocks the other, resulting in false alarm.

[0076] The embodiment provides a post-processing method for false alarm filtering of behavior wearing in a smart factory, as shown in FIG. 9, which comprises the following steps: Figure 1

[0077] S1, acquiring historical monitoring images of a factory area, and constructing an image data set of employee behavior wearing; specifically, extracting frame images of historical videos of the factory area scene, manually labeling the categories of each image, the categories including "human body", "head" and wearing requirements of the smart factory, and taking the labeled images as the image data set.

[0078] The wearing requirements of the smart factory include "wearing a safety helmet", "wearing a tool", "wearing a coat", "wearing ordinary shoes" and "wearing anti-throw shoes", and the labeled images are taken as the image data.

[0079] S2, constructing a cascade-rcnn detection network model for employee behavior wearing behavior, and inputting the constructed image data set into the cascade-rcnn detection network model for training.

[0080] S3, using the trained cascade-rcnn detection network model to intelligently detect the behavior wearing state of employees in the factory area, and the detection result includes a "human body" target frame and target information thereof, a "head class" target frame and target information thereof, and the "head class" target frame includes a "head" target frame and a "wearing a safety helmet" target frame.

[0081] ​The target information includes: target bounding box coordinates and class confidence. The target bounding box coordinates include the coordinates of the top-left corner and the bottom-right corner. Filtering by the top-left and bottom-right corners of the bounding box does not require large-scale modifications and improvements to the model structure, nor does it require retraining or replacing the model. This results in low practical application costs and wide applicability. The class confidence is the confidence that the target bounding box belongs to the currently detected class.

[0082] The intelligent detection yields a list of "human body" categories: [human body 1, human body 2, human body 3...];

[0083] List of "head" categories [head1, head2, head3, ...];

[0084] List of categories for "Helmet Wearing": [Helmet Wearing 1, Helmet Wearing 2, Helmet Wearing 3, ...];

[0085] And the corresponding confidence list of the bounding boxes: [confidence 1, confidence 2, confidence 3, ...];

[0086] The detection results for each target are as follows:

[0087] {"coordinate":[x min ,y min ,x max ,y max ],

[0088] "score":score,

[0089] "tagName":labels

[0090] }

[0091] Where labels are category names, x min ,y min x represents the coordinates of the top-left corner of the target bounding box predicted by the model. max ,y max The coordinates of the bottom right corner of the target bounding box predicted by the model are given, and the score is the confidence score of the predicted class.

[0092] S4. Analyze the detection results of the cascade-rcnn detection network model to obtain the cross ratio between the "human" target box and the "head" target box, set the filtering threshold, and perform the first filtering.

[0093] If the cross ratio is greater than the filtering threshold, the "head" target box corresponding to the "human body" target box is retained, that is, the "head" class is considered to belong to the current "human body".

[0094] If one "human body" target frame corresponds to one "head class" target frame, and the "head class" target frame is a "safety helmet" target frame, the verification ends, that is, the "human body" wears a safety helmet; if the "head class" target frame is a "head" target frame, an alarm of wearing an illegal safety helmet is pushed, that is, the "human body" does not wear a safety helmet;

[0095] If one "human body" target frame corresponds to multiple "head class" target frames, that is, the "human body" may have multiple heads, secondary filtering is performed;

[0096] The S4 analyzes the detection result of the cascade-rcnn detection network model to obtain the intersection ratio between the "human body" target frame and the "head class" target frame, and the specific process is as follows:

[0097] All "head" and "safety helmet" target frames in the current "human body" target frame are obtained, and the next step of filtering is performed;

[0098] The intersection ratio is calculated by the following formula:

[0099] The left upper corner point A coordinate (x min ,y min ) and the right lower corner point B coordinate (x max ,y max ) of the intersection part of the "human body" target frame and the "head class" target frame are calculated: x min =max(box1[0],box2[0]);y min =max(box1[1],box2[1]);x max =min(box1[2],box2[2]);y max =min(box1[3],box2[3]);

[0100] Wherein, the box1 refers to the coordinate list of the "human body" target frame, including; [x1 min ,y1 min ,x1 max ,y1 max ], wherein box1[i] is the ith element in box1, i=0, 1, 2, 3; the box2 refers to the coordinate list of the "head class" target frame, including; [x2 min ,y2 min ,x2 max ,y2 max ], wherein box2[i] is the ith element in box2, i=0, 1, 2, 3;

[0101] For example, box1[0] represents the first value x1 minbox1[1] represents the second value y1 of the "human body" target frame coordinate list min box1[2] represents the third value x1 of the "human body" target frame coordinate list max box1[3] represents the fourth value y1 of the "human body" target frame coordinate list max The upper left corner point coordinates of the "human body" target frame are (x1 min , y1 min ), and the right lower corner point coordinates of the "human body" target frame are (x1 max , y1 max ); box2[0] represents the first value x2 of the "head class" target frame coordinate list min box2[1] represents the second value y2 of the "head class" target frame coordinate list min box2[2] represents the third value x2 of the "head class" target frame coordinate list max box2[3] represents the fourth value y2 of the "head class" target frame coordinate list max The upper left corner point coordinates of the "head class" target frame are (x2 min , y2 min ), and the right lower corner point coordinates of the "head class" target frame are (x2 max , y2 max );

[0102] Calculate the area of the frame of the intersection part:

[0103] intersection = max(0, x max -x min )*max(0, y max -y min ) (1);

[0104] Calculate the intersection ratio, i.e. the ratio of the frame of the intersection part to the "human body" target frame:

[0105] box1_area = (box1[2]-box1[0]+1)*(box1[3]-box1[1]+1) (2);

[0106] cover_percentage = float(intersection_area) / float(box1_area) (3);

[0107] Wherein, box1_area refers to the area of the "human body" target frame, cover_percentage refers to the intersection ratio, and float refers to a decimal.

[0108] S5, when one "human body" target frame corresponds to multiple "head" target frames, that is, the "human body" may have multiple heads, then secondary filtering is performed:

[0109] All "human body" target frames are traversed, the distance from the top left corner point of the "human body" target frame to the top left corner point of the corresponding "head" target frame is calculated, and the distance from the middle point of the top of the "human body" target frame to the middle point of the top of the corresponding "head" target frame is calculated, and the average distance distance is calculated and a distance list [distance1, distance2, distance3,...] is formed. The specific is as follows:

[0110] The distance between the two target frames is calculated as follows:

[0111] The obtained target frame coordinate values include the coordinate values of the left top corner point and the right bottom corner point of the target frame;

[0112] The left top corner coordinate (xl min , yl min ) and the right bottom corner coordinate (xl max , yl max ) of the "human body" target frame, and the left top corner coordinate (x2 min , y2 min ) and the right bottom corner coordinate (x2 max , y2 max ) of the "head" target frame are obtained. The final distance distance between the "human body" target frame and the "head" target frame is obtained through the following formula:

[0113] ;

[0114] ;

[0115] (4);

[0116] (5);

[0117] The left line distance from the "human body" target frame to the multiple "head" target frames corresponding to the "human body" target frame is calculated:

[0118] (6);

[0119] The middle line distance from the "human body" target frame to the multiple "head" target frames corresponding to the "human body" target frame is calculated:

[0120] (7);

[0121] The average of the left line distance and the middle line distance is calculated:

[0122] (8);

[0123] wherein, x1 ul is the horizontal coordinate value of the top middle point of the first target frame, x2 ul is the horizontal coordinate value of the top middle point of the second target frame, x1 um is the horizontal coordinate value of the top middle point of the first target frame, x2 um is the horizontal coordinate value of the top middle point of the second target frame.

[0124] The first target frame is a "human body" target frame, and the second target frame is a "head" target frame.

[0125] S6, judging whether to push the alarm of wearing non-compliance according to the mean value of the distance 1 and the distance 2 and the category confidence:

[0126] Obtaining a category confidence list of all "head" and "head safety helmet" category corresponding target frames corresponding to the "human body" target frame, and arranging in order;

[0127] Case one: if the "head" target frame with the minimum distance to the "human body" target frame and the "head" target frame with the highest category confidence are the same, then the "head" target frame is retained;

[0128] Case two: if the "head" target frame with the minimum distance to the "human body" target frame and the "head" target frame with the highest category confidence are not the same, then the category confidence difference value between the "head" target frame with the minimum distance to the "human body" target frame and other "head" target frames is calculated and the absolute value is taken, a preset difference value threshold is set, if the absolute value of the difference value is less than the difference value threshold, then the "head" target frame with the minimum distance to the "human body" target frame is retained, otherwise, the "head" target frame with the highest category confidence is retained;

[0129] If the "head" target frame exists in the retained "head" target frame, the alarm of wearing non-compliance is pushed (the "head" target frame is pushed to the monitoring system, and the alarm of wearing non-compliance is given).

[0130] Obviously, the above embodiments of the present application are only examples for clearly illustrating the technical solutions of the present application, and are not intended to limit the specific embodiments of the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A post-processing method for smart factory behavior wear mis-advice alarm filtering, characterized in that, Comprise; S1, acquire factory history monitoring image, construct employee behavior wearing image data set; S2, construct cascade-rcnn detection network model for employee behavior wearing behavior, and input the constructed image data set into cascade-rcnn detection network model training; S3, the wearing state of employee behavior in factory area is intelligently detected by using the trained cascade-rcnn detection network model;The detection result includes "human" target frame and its target information, "head class" target frame and its target information, and the "head class" target frame includes "head" target frame and "safety helmet" target frame; The target information includes target frame coordinate value and category confidence, and the target frame coordinate value includes left upper corner point coordinate value and right lower corner point coordinate value, which are filtered by the left upper corner point and right lower corner point of the coordinate frame; S4, first filtering: The detection result of cascade-rcnn detection network model is analyzed, and the first filtering is carried out, the intersection ratio between "human" target frame and "head class" target frame is obtained, and the filtering threshold is set: If the intersection ratio is greater than the filtering threshold, the "head class" target frame corresponding to the "human" target frame is retained;If the intersection ratio is less than or equal to the filtering threshold, the "head class" target frame corresponding to the "human" target frame is filtered; S5, second filtering: When one "human" target frame corresponds to multiple "head class" target frames, second filtering is carried out; All "human" target frames are traversed, the left line distance from "human" target frame to multiple "head class" target frames corresponding to the "human" target frame is calculated, and the middle line distance from "human" target frame to multiple "head class" target frames corresponding to the "human" target frame is calculated, the mean value of left line distance and middle line distance is calculated and distance list is formed; The left line distance is the straight line distance from the left upper corner point of "human" target frame to the left upper corner point of "head class" target frame; The middle line distance is the straight line distance from the middle point of the top of "human" target frame to the middle point of the top of "head class" target frame; S6, according to the distance list and the category confidence, it is judged whether to push the alarm of wearing non-compliance: The category confidence list of all "head class" target frames corresponding to "human" target frame is obtained; Case one: if the "head class" target frame closest to "human" target frame and the "head class" target frame with the highest category confidence are the same, the "head class" target frame is retained; Case two: if the "head class" target frame closest to "human" target frame and the "head class" target frame with the highest category confidence are not the same, the category confidence difference between the "head class" target frame closest to "human" target frame and other "head class" target frames is calculated and the absolute value is taken, and a preset difference threshold is set; If the absolute value of the category confidence difference is less than the difference threshold, the "head class" target frame closest to "human" target frame is retained;Otherwise, the "head class" target frame with the highest category confidence is retained; If there is "head" target frame in the retained "head class" target frame, the alarm of wearing non-compliance is pushed.

2. The post-processing method of false alarm filtering of behavior wear in a smart factory according to claim 1, characterized in that, The S1 specifically comprises: extracting frame images of a plant area scene historical video, manually labeling categories of each image, the categories comprising "human body", "head", and wearing requirements of the smart factory, and taking the labeled images as an image data set.

3. The post-processing method of false alarm filtering of behavior wear in a smart factory according to claim 2, characterized in that, The wearing requirements of the smart factory comprise "wearing a safety helmet", "wearing a tool", "wearing a coat", "wearing ordinary shoes", and "wearing anti-throw shoes".

4. The post-processing method of false alarm filtering of behavior wear in a smart factory according to claim 2, characterized in that, The S4 analyzes a detection result of the cascade-rcnn detection network model to obtain a cross ratio between a "human body" target frame and a "head" target frame, specifically as follows: All "head" and "safety helmet" target frames in the current "human body" target frame are obtained, and then a next step of filtering is performed. The cross ratio is calculated by the following formula: Calculate the intersection part of the "human" target frame and the "head class" target frame left upper corner point A coordinate (x min ,y min ) and right lower corner point B coordinate (x max ,y max ): x min =max(box1[0],box2[0]); y min =max(box1[1],box2[1]); x max =min(box1[2],box2[2]); y max =min(box1[3],box2[3]); The box1 refers to the coordinate list of the "human body" target frame, including; [x1 min ,y1 min ,x1 max ,y1 max ], wherein box1[i] is the i-th element in the box1, i=0, 1, 2, 3; the box2 refers to the coordinate list of the "head class" target frame, including; [x2 min ,y2 min ,x2 max ,y2 max ], wherein box2[i] is the i-th element in the box2, i=0, 1, 2, 3; An area of the frame of the intersection part is calculated: intersection = max(0, x max -x min ) * max(0, y max -y min ) (1); The cross ratio, that is, a ratio of the frame of the intersection part to the "human body" target frame, is calculated: box1_area=(box1[2]-box1[0]+1)*(box1[3]-box1[1]+1) (2); cover_percentage=float(intersection_area) / float(box1_area) (3); wherein box1_area refers to an area of the "human body" target frame, cover_percentage refers to the cross ratio, and float refers to taking a decimal.

5. The post-processing method of false alarm filtering of behavior wear in a smart factory according to claim 4, characterized in that, The calculation manner of the S5 is specifically as follows: Obtained target frame coordinate values comprise coordinate values of a left upper corner point of a target frame and a right lower corner point of the target frame; Calculate the top left corner of the "human body" target bounding box (xl min yl min (x2) to the top left corner of the corresponding "header" target box min y2 min The distance on the left side of the line. ， And the distance between the top center point of the "human body" target box and the top center point of the "head" target box. ，， ; And a mean value distance of the left line distance and the middle line distance is calculated. ； ； （4）； （5）； （6）； （7）； （8）； wherein x1 ul is a horizontal coordinate value of the top middle point of the first target frame, x2 ul is a horizontal coordinate value of the top middle point of the second target frame, x1 um is a horizontal coordinate value of the top middle point of the first target frame, x2 um is a horizontal coordinate value of the top middle point of the second target frame. The first target frame is the "human body" target frame, and the second target frame is the "head" target frame.

Images