CN115393905A - Helmet wearing detection method based on attitude correction - Google Patents

Helmet wearing detection method based on attitude correction Download PDF

Info

Publication number
CN115393905A
CN115393905A CN202211356734.4A CN202211356734A CN115393905A CN 115393905 A CN115393905 A CN 115393905A CN 202211356734 A CN202211356734 A CN 202211356734A CN 115393905 A CN115393905 A CN 115393905A
Authority
CN
China
Prior art keywords
head
human head
helmet
image
human
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211356734.4A
Other languages
Chinese (zh)
Inventor
康凯
艾坤
刘海峰
王子磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hefei Zhongke Leinao Intelligent Technology Co ltd
Original Assignee
Hefei Zhongke Leinao Intelligent Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hefei Zhongke Leinao Intelligent Technology Co ltd filed Critical Hefei Zhongke Leinao Intelligent Technology Co ltd
Priority to CN202211356734.4A priority Critical patent/CN115393905A/en
Publication of CN115393905A publication Critical patent/CN115393905A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/24—Aligning, centring, orientation detection or correction of the image
    • G06V10/242—Aligning, centring, orientation detection or correction of the image by image rotation, e.g. by 90 degrees
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Multimedia (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Human Computer Interaction (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a helmet wearing detection method based on posture correction, and belongs to the technical field of safety appliance wearing detection. The invention combines the characteristics of the human head target, adds the human head key point detection branch for the human head detector, the added branch is compatible with the current mainstream detector, and the detection precision of the human head can be improved by introducing extra supervision information; the posture of the human head is corrected through the detected key points of the human head, so that the influence of the human head posture on human head attribute classifiers such as safety helmet wearing and the like can be weakened, and the classification precision is improved; when the head attribute classifier such as wearing of a safety helmet and the like is trained, random shaking of key points of the head is adopted for data enhancement, and the robustness of the head attribute classifier can be improved.

Description

Helmet wearing detection method based on attitude correction
Technical Field
The invention relates to the technical field of safety appliance wearing detection, in particular to a safety helmet wearing detection method based on posture correction.
Background
The safety helmet is an indispensable safety appliance for safety production workers in all walks of life. However, due to the intention or the accident of the working personnel, the situation that the safety helmet is not worn frequently occurs in the actual operation, and certain potential safety hazard is caused. In order to guarantee the personal safety of workers, the traditional mode is mainly to monitor or watch monitoring videos and give an early warning through safety supervision personnel on site, but the mode is low in efficiency, and the safety supervision personnel are easy to fatigue and prone to false detection.
For example, in the prior art, the human head target detection is performed by using a self-universal target detection algorithm for reference, and the algorithm design is not performed in combination with the characteristics of the human head, so that the human head posture and the visual angle of a camera in a real scene are complicated and changeable, and the influence of the human head posture on the safety helmet wearing detection is not considered by the current safety helmet wearing attribute classifier. Therefore, a helmet wearing detection method based on posture correction is provided.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: how to design a detection algorithm by combining the characteristics of a human head target, improve the detection precision of the human head target, weaken the influence of human head posture on a human head attribute classifier such as safety helmet wearing and the like, improve the classification precision of human head attributes, and provide a safety helmet wearing detection method based on posture correction.
The invention solves the technical problems through the following technical scheme, and comprises the following steps:
s1: scaling an input image to a set size, sending the input image to a head detector, and detecting to obtain a head bounding box and head key points;
s2: for each detected human head target, intercepting a human head image from an input image according to the detected human head bounding box; calculating an alignment transformation matrix according to the corresponding relation between the detected head key points and the standard head key points, aligning the head image to the standard posture and size by using the alignment transformation matrix, and enabling the target image to be an aligned head image; sending the aligned head images into a head attribute classifier, and outputting to obtain multiple head attributes; and judging whether the safety helmet is worn or not by using the output head attributes.
Further, in step S1, the head detector includes a first backbone network module, a detection head module, and a post-processing module, where the detection head module includes three branches, which are a head classification branch, a head detection branch, and a head key point detection branch, an input image first enters the first backbone network module to obtain image feature representations on multiple spatial scales, and features of each scale are then sent to three branches with independent parameters in the detection head module to obtain a head confidence feature map, a head bounding box feature map, and a head key point coordinate feature map, and the head confidence feature map, the head bounding box feature map, and the head key point coordinate feature map are processed by the post-processing module to obtain final head bounding box and head key point coordinates.
Further, in training the human head detector, each branch leads out a loss term, and for a single sample, the mathematical form of the final loss function is:
Figure 274332DEST_PATH_IMAGE001
wherein the content of the first and second substances,
Figure 616452DEST_PATH_IMAGE002
is a head classification loss item which adopts a softmax cross entropy loss function,
Figure 934301DEST_PATH_IMAGE003
is the probability of being predicted as a human head,
Figure 652858DEST_PATH_IMAGE004
is a true tag of the target;
Figure 104044DEST_PATH_IMAGE005
is a regression loss term of the head bounding box, the loss term adopts a Smooth-L1 loss function,
Figure 882644DEST_PATH_IMAGE006
is the predicted head bounding box of the person,
Figure 422210DEST_PATH_IMAGE007
is a real head bounding box;
Figure 944458DEST_PATH_IMAGE008
is a human head key point regression loss term which adopts a Smooth-L1 loss function,
Figure 253079DEST_PATH_IMAGE009
is a key point of the head of a person to be predicted,
Figure 202581DEST_PATH_IMAGE010
is a real head key point;
Figure 963863DEST_PATH_IMAGE011
，
Figure 555382DEST_PATH_IMAGE012
the loss term is only effective for human head targets, in addition by
Figure 718510DEST_PATH_IMAGE013
And
Figure 838913DEST_PATH_IMAGE014
to control the weight occupied between the lossy terms.
Further, in the step S2, after the head bounding box is detected and enlarged by 1.3 times, the head image corresponding to the head bounding box is cut from the input image.
Further, in the step S2, let
Figure 353071DEST_PATH_IMAGE015
Is the key point of the detected human head,
Figure 217121DEST_PATH_IMAGE016
is a key point of the corresponding standard head of a person,
Figure 500335DEST_PATH_IMAGE017
the relationship between the detected head keypoints and the standard head keypoints is modeled using a similarity transformation, as follows:
Figure 57218DEST_PATH_IMAGE018
wherein the content of the first and second substances,
Figure 793093DEST_PATH_IMAGE019
is a matrix of similarity transformation, which is,
Figure 195256DEST_PATH_IMAGE020
is the translation parameter(s) of the image,
Figure 332976DEST_PATH_IMAGE021
as a function of the scale parameter(s),
Figure 57831DEST_PATH_IMAGE022
is a rotation parameter;
note the book
Figure 15423DEST_PATH_IMAGE023
By using
Figure 752434DEST_PATH_IMAGE024
The corresponding points are listed in the following equation:
Figure 744661DEST_PATH_IMAGE025
；
denote the above formula as
Figure 377768DEST_PATH_IMAGE026
Wherein
Figure 88235DEST_PATH_IMAGE027
The equation is an over-determined equation with a least-squares solution of
Figure 97779DEST_PATH_IMAGE028
(ii) a According to the obtained
Figure 210092DEST_PATH_IMAGE029
A similarity transformation matrix can be obtained
Figure 279679DEST_PATH_IMAGE030
(ii) a Transforming the matrix
Figure 211863DEST_PATH_IMAGE030
And acting on the human head image to obtain the aligned human head image.
Further, in step S2, the head attribute classifier includes a second backbone network module and a plurality of classification branches, the plurality of classification branches are respectively connected to the second backbone network module, the input image enters the second backbone network module to obtain a feature representation of the image, and then the input image is sent to the branches with independent parameters to obtain a plurality of head attributes corresponding to the plurality of classification branches.
Furthermore, the number of the classification branches is at least two, the classification branches are respectively a helmet wearing detection branch and a helmet color classification branch, and the helmet wearing confidence level and the confidence level of each color of the helmet are correspondingly output.
Further, a multi-tasking loss function is used in training the head attribute classifier, and for a single sample, each branch leads out a loss term, and the mathematical form of the loss function is as follows:
Figure 25098DEST_PATH_IMAGE031
wherein the content of the first and second substances,
Figure 257496DEST_PATH_IMAGE032
is a safety helmet wearing partA class loss term, which employs a softmax cross-entropy loss function,
Figure 232406DEST_PATH_IMAGE033
is the predicted probability of headgear being worn,
Figure 917465DEST_PATH_IMAGE034
is a real tag worn by the safety helmet;
Figure 268812DEST_PATH_IMAGE035
is a safety helmet color classification loss item, the loss item adopts a softmax cross entropy loss function,
Figure 355716DEST_PATH_IMAGE036
is the predicted color of the helmet,
Figure 501527DEST_PATH_IMAGE037
is a real helmet color tag;
Figure 408303DEST_PATH_IMAGE038
the loss term is only effective for the head of the person wearing the safety helmet.
Furthermore, when training the head attribute classifier, random jitter of key points of the head is introduced for data enhancement, and the specific operations are as follows: inputting a human head image and key points thereof, adding random offset to the human head key points, namely random jitter, and then aligning the human head image.
Further, in the step S2, the method for determining whether or not to wear the crash helmet using the output attribute of the head of the person in the closed construction site scene includes the steps of:
a1: presetting a threshold;
a2: and when the wearing confidence of the safety helmet of the head image is larger than the threshold value, judging that the person wears the safety helmet, otherwise, judging that the person does not wear the safety helmet.
Compared with the prior art, the invention has the following advantages: the helmet wearing detection method based on the posture correction combines the characteristics of a human head target, adds a human head key point detection branch for a human head detector, the added branch is compatible with a current mainstream detector, and the detection precision of the human head can be improved by introducing extra supervision information; the gesture of the human head is corrected through the detected key points of the human head, so that the influence of the human head gesture on a head attribute classifier such as helmet wearing and the like can be weakened, and the classification precision is improved; when the head attribute classifier such as wearing of a safety helmet and the like is trained, random shaking of key points of the head is adopted for data enhancement, and the robustness of the head attribute classifier can be improved.
Drawings
FIG. 1 is a schematic flow chart of a method for detecting wearing of a helmet based on posture correction according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a human head detector according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a head attribute classifier according to an embodiment of the present invention;
FIG. 4 is a schematic view of a wearing manner of the safety helmet according to an embodiment of the present invention;
FIG. 5 is a second schematic view of a manner of determining how to wear a safety helmet according to an embodiment of the present invention.
Detailed Description
The following examples are given for the detailed implementation and specific operation of the present invention, but the scope of the present invention is not limited to the following examples.
As shown in fig. 1, the present embodiment provides a technical solution: a helmet wearing detection method based on posture correction comprises the following steps:
1) The input image is scaled to a fixed size and sent to a head detector, while head bounding boxes and head keypoints are detected.
2) For each human head target detected, the following operations are performed:
2.1 Intercept the human head image from the input image according to the detected human head bounding box;
2.2 According to the corresponding relation between the detected head key points and the standard head key points, an alignment transformation matrix is calculated, the head images are aligned to the standard posture and size by using the alignment transformation matrix, and the target images become aligned head images;
2.3 Sending the aligned head images into a head attribute classifier, and obtaining head attributes such as the confidence of wearing safety helmets and the confidence of safety helmets of all colors;
2.4 The above output and other information (i.e., the additional information according to the present invention) are used to determine whether or not to wear the crash helmet.
In the step S1:
as shown in fig. 2, an input image first enters a backbone network module to obtain image feature representations on a plurality of spatial scales (fig. 2 only shows 3 scales), and features of each scale are sent to three independent branches of a parameter in a detection head module: the head classification branch, the head detection branch and the head key point detection branch are respectively responsible for predicting the head confidence coefficient, the head bounding box and the coordinates of the head key point. The result of each branch is processed by a post-processing module (mainly non-maximum suppression) to obtain the final head bounding box and the head key point coordinates.
The backbone Network module is realized by adopting convolutional neural networks such as ResNet, denseNet, darkNet or MobileNet, and FPN (Feature Pyramid Network) is added on the backbone Network module to perform multi-scale Feature fusion so as to improve the expression capability of features; each branch is also a small convolutional neural network, assuming that their input feature map size is
Figure 560411DEST_PATH_IMAGE039
The output feature size of the head classification branch is
Figure 236243DEST_PATH_IMAGE040
(the last dimension refers to the probability of a human head and the probability of a non-human head), the output feature map size of the human head detection branch is
Figure 818534DEST_PATH_IMAGE041
(the last dimension refers to the upper left point coordinate of the bounding box and the bounding box width and height, orTheir offsets relative to the corresponding anchor frame), the output feature map size of the human head keypoint detection branch is
Figure 478186DEST_PATH_IMAGE042
(the last dimension refers to the coordinates of the four head keypoints, or their offset relative to the corresponding anchor frame).
In training the human head detector, each branch leads out a loss term, and the mathematical form of the final loss function (for a single sample) is:
Figure 171335DEST_PATH_IMAGE001
wherein the content of the first and second substances,
Figure 967253DEST_PATH_IMAGE002
is a head classification loss term of the person,
Figure 720445DEST_PATH_IMAGE003
is the probability of being predicted as a human head,
Figure 336234DEST_PATH_IMAGE004
if the target is a real label (if the human head label is 1, the non-human head label is 0), the loss item adopts a softmax cross entropy loss function;
Figure 98654DEST_PATH_IMAGE005
is a regression loss term of the head bounding box,
Figure 749078DEST_PATH_IMAGE006
is the predicted head bounding box of the person,
Figure 673172DEST_PATH_IMAGE007
the loss item is a real human head bounding box and adopts a Smooth-L1 loss function;
Figure 41836DEST_PATH_IMAGE008
is a regression loss term of key points of the human head,
Figure 76788DEST_PATH_IMAGE009
is a key point of the head of a person to be predicted,
Figure 847298DEST_PATH_IMAGE010
is a real head key point; the loss term also employs the Smooth-L1 loss function.
Figure 942293DEST_PATH_IMAGE011
，
Figure 532675DEST_PATH_IMAGE012
The loss term is only effective for the human head target, so it is preceded by
Figure 636897DEST_PATH_IMAGE004
Coefficient of otherwise passing
Figure 261913DEST_PATH_IMAGE013
And
Figure 524880DEST_PATH_IMAGE014
to control the weight occupied between the lossy terms.
In step 2.1:
considering that the detected human head image is rotated by the subsequent alignment operation, in order to avoid introducing too many invalid regions on the rotated human head image, the detected human head bounding box is enlarged by 1.3 times, and then the human head image is intercepted according to the enlarged human head bounding box.
In step 2.2:
is provided with
Figure 602557DEST_PATH_IMAGE015
Is a key point of the detected human head,
Figure 244891DEST_PATH_IMAGE016
is a key point of the corresponding standard head of a person,
Figure 724414DEST_PATH_IMAGE017
(ii) a Standard headThe key point is the calculated mean over one additional data set.
In this embodiment, four standard head key points of the head, the neck, the left ear and the right ear are adopted, and the relationship between the detected head key points and the standard head key points is modeled by similarity transformation, as follows:
Figure 161212DEST_PATH_IMAGE018
wherein the content of the first and second substances,
Figure 726185DEST_PATH_IMAGE019
is a matrix of similarity transformation, which is,
Figure 906631DEST_PATH_IMAGE020
is a parameter of the translation that is,
Figure 506239DEST_PATH_IMAGE021
in order to be a scale parameter,
Figure 113938DEST_PATH_IMAGE044
is a rotation parameter;
note the book
Figure 166208DEST_PATH_IMAGE046
By using
Figure 150344DEST_PATH_IMAGE024
The corresponding points are listed in the following equation:
Figure 604459DEST_PATH_IMAGE047
denote the above formula as
Figure 383060DEST_PATH_IMAGE026
Wherein
Figure 657046DEST_PATH_IMAGE027
Because of
Figure 179294DEST_PATH_IMAGE048
So the equation is generally an over-determined equation with a least squares solution of
Figure 753495DEST_PATH_IMAGE028
. According to the obtained
Figure 702997DEST_PATH_IMAGE029
A similarity transformation matrix can be obtained
Figure 62866DEST_PATH_IMAGE030
(ii) a Transforming the matrix
Figure 123226DEST_PATH_IMAGE030
And acting on the head image to obtain the aligned head image.
In step 2.3:
as shown in fig. 3, an input image firstly enters a backbone network module to obtain a feature representation of the image, and then is sent to a branch network with independent parameters, such as a helmet wearing detection branch and a helmet color classification branch, to respectively predict a helmet wearing confidence and confidences of colors of a helmet; other branches (such as head pose, occlusion degree, age, etc.) can also be added according to the application scene of the head attribute classifier.
The backbone network can be implemented by a classical CNN network, such as ResNet, densnet, mobileNet, and the like. Each branch leads out a loss item, and a multitask loss function is adopted when the head attribute classifier is trained, and the mathematical form (for a single sample) of the loss item is as follows:
Figure 551933DEST_PATH_IMAGE031
wherein the content of the first and second substances,
Figure 672336DEST_PATH_IMAGE032
is a classification loss item for wearing the safety helmet,
Figure 920915DEST_PATH_IMAGE033
is the predicted probability of headgear wear,
Figure 50545DEST_PATH_IMAGE034
the loss term is a real label worn by the safety helmet (the label with the safety helmet is 1, and the label without the safety helmet is 0), and the loss term adopts a softmax cross entropy loss function.
Figure 68180DEST_PATH_IMAGE035
Is a color classification loss term of the safety helmet,
Figure 359484DEST_PATH_IMAGE036
is the predicted color of the helmet,
Figure 95358DEST_PATH_IMAGE037
is a true helmet color label, and the loss term also adopts a softmax cross entropy loss function.
Figure 763100DEST_PATH_IMAGE038
The loss item is only effective for the head wearing the safety helmet, so that the front part of the safety helmet is provided with
Figure 166400DEST_PATH_IMAGE034
And (4) the coefficient.
It should be noted that, when training the head attribute classifier, in order to weaken the influence caused by the detection error of the head key point and improve the robustness of the model, random jitter data enhancement of the head key point is introduced, and the specific operations are as follows: inputting the human head image and the key points thereof, adding random offset (namely random jitter) to the human head key points, and then carrying out the alignment operation on the human head image. The data enhancement mode can be carried out off-line or on-line.
In step 2.4:
and judging whether the safety helmet is worn by using the output of the head attribute classifier and other information.
Mode one (see fig. 4): and presetting a threshold, when the wearing confidence of the safety helmet of the head image is greater than the threshold, judging that the person wears the safety helmet, otherwise, judging that the person does not wear the safety helmet. The method is suitable for closed construction site scenes.
Mode two (see fig. 5): firstly, a worker judging module (mainly comprising a human body detector and a human body attribute classifier) is utilized to judge whether a detected person is a worker, the worker is judged according to a first mode, and is directly ignored in a second non-worker mode, and the method is suitable for a scene that passers-by often pass around a construction site.
It should be noted that the head attribute such as wearing a helmet is independent of the head pose angle, but for the head attribute classifier, the pose angle of the head target in the input image may be changed, and different prediction results may be obtained, that is, the head attribute classifier does not have rotation invariance. In order to weaken the influence of the head postures on the head attribute classifier, the invention introduces the alignment of the key points of the head, and the operation can align the head images in different postures to the standard size and turn right. In addition, through the alignment of the key points of the human head, the background area in the human head image can be reduced, and the dependence of the human head attribute classifier on a specific human head detector is weakened (because the bounding boxes detected by different human head detectors are different in size, the standard size with the same normalized style of the human head image can be obtained through the alignment of the key points of the human head).
To sum up, the helmet wearing detection method based on posture correction of the above embodiment combines the characteristics of the human head target, adds the human head key point detection branch to the human head detector, the added branch is compatible with the current mainstream detector, and the detection precision of the human head can be improved by introducing extra supervision information; the posture of the human head is corrected through the detected key points of the human head, so that the influence of the human head posture on human head attribute classifiers such as safety helmet wearing and the like can be weakened, and the classification precision is improved; when the head attribute classifier such as wearing of a safety helmet and the like is trained, random shaking of key points of the head is adopted for data enhancement, and the robustness of the head attribute classifier can be improved.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims (8)

1. A helmet wearing detection method based on posture correction is characterized by comprising the following steps:
s1: the input image is zoomed to a set size and sent to a human head detector, and a human head bounding box and human head key points are obtained through detection;
s2: for each detected human head target, intercepting a human head image from an input image according to the detected human head bounding box; calculating an alignment transformation matrix according to the corresponding relation between the detected head key points and the standard head key points, aligning the head image to the standard posture and size by using the alignment transformation matrix, and enabling the target image to be an aligned head image; sending the aligned head images into a head attribute classifier, and outputting to obtain multiple head attributes; and judging whether the safety helmet is worn or not by using the output head attributes.
2. The helmet wearing detection method based on posture correction according to claim 1, characterized in that: in step S1, the head detector includes a first backbone network module, a detection head module, and a post-processing module, where the detection head module includes three branches, which are a head classification branch, a head detection branch, and a head key point detection branch, an input image first enters the first backbone network module to obtain image feature representations on multiple spatial scales, and features of each scale are sent to three branches with independent parameters in the detection head module to obtain a head confidence feature map, a head bounding box feature map, and a head key point coordinate feature map, respectively, and the head confidence feature map, the head bounding box feature map, and the head key point coordinate feature map are processed by the post-processing module to obtain final head bounding box and head key point coordinates.
3. The helmet wearing detection method based on posture correction according to claim 2, characterized in that: in training the human head detector, each branch leads out a loss term, and for a single sample, the mathematical form of the final loss function is:
Figure 693812DEST_PATH_IMAGE001
wherein the content of the first and second substances,
Figure 378872DEST_PATH_IMAGE002
is a head classification loss item which adopts a softmax cross entropy loss function,
Figure 730218DEST_PATH_IMAGE003
is the probability of being predicted as a human head,
Figure 551544DEST_PATH_IMAGE004
is the true tag of the target;
Figure 962934DEST_PATH_IMAGE005
is a regression loss term of the head bounding box, the loss term adopts a Smooth-L1 loss function,
Figure 869710DEST_PATH_IMAGE006
is the predicted head bounding box of the person,
Figure 759168DEST_PATH_IMAGE007
is a real head bounding box;
Figure 435000DEST_PATH_IMAGE008
is a human head key point regression loss term which adopts a Smooth-L1 loss function,
Figure 751712DEST_PATH_IMAGE009
is a key point of the head of a person to be predicted,
Figure 145784DEST_PATH_IMAGE010
is a real head key point;
Figure 635672DEST_PATH_IMAGE011
，
Figure 431589DEST_PATH_IMAGE012
the loss term is only effective for human head targets, in addition by
Figure 184782DEST_PATH_IMAGE013
And
Figure 800571DEST_PATH_IMAGE014
to control the weight occupied between the lossy terms.
4. The helmet wearing detection method based on posture correction according to claim 3, characterized in that: in the step S2, let
Figure 31832DEST_PATH_IMAGE015
Is a key point of the detected human head,
Figure 682256DEST_PATH_IMAGE016
is the key point of the corresponding standard head,
Figure 403087DEST_PATH_IMAGE017
the relationship between the detected head keypoints and the standard head keypoints is modeled using a similarity transformation, as follows:
Figure 771752DEST_PATH_IMAGE018
wherein the content of the first and second substances,
Figure 803774DEST_PATH_IMAGE019
is a matrix of similarity transformation, which is,
Figure 308705DEST_PATH_IMAGE020
is the translation parameter(s) of the image,
Figure 403700DEST_PATH_IMAGE021
as a function of the scale parameter(s),
Figure 994081DEST_PATH_IMAGE023
is a rotation parameter;
note the book
Figure 567145DEST_PATH_IMAGE025
By using
Figure 988899DEST_PATH_IMAGE026
The corresponding points are listed in the following equation:
Figure 254795DEST_PATH_IMAGE027
；
denote the above formula as
Figure 332473DEST_PATH_IMAGE028
Wherein
Figure 974807DEST_PATH_IMAGE029
The equation is an overdetermined equation having a least squares solution of
Figure 188750DEST_PATH_IMAGE030
(ii) a According to the obtained
Figure 625548DEST_PATH_IMAGE031
A similarity transformation matrix can be obtained
Figure 190522DEST_PATH_IMAGE032
(ii) a Transforming the matrix
Figure 105388DEST_PATH_IMAGE032
And acting on the head image to obtain the aligned head image.
5. The helmet wearing detection method based on posture correction according to claim 4, characterized in that: in step S2, the head attribute classifier includes a second backbone network module and a plurality of classification branches, the plurality of classification branches are respectively connected to the second backbone network module, the input image enters the second backbone network module first to obtain the feature representation of the image, and then the input image is sent to the branches with independent parameters to obtain a plurality of head attributes corresponding to the plurality of classification branches.
6. The helmet wearing detection method based on posture correction according to claim 5, characterized in that: the number of the classification branches is at least two, the classification branches are respectively a helmet wearing detection branch and a helmet color classification branch, and the helmet wearing confidence coefficient and the confidence coefficient of each color of the helmet are correspondingly output.
7. The helmet wearing detection method based on posture correction according to claim 5, characterized in that: when the head attribute classifier is trained, random shaking of key points of the head is introduced for data enhancement, and the specific operation is as follows: inputting a human head image and key points thereof, adding random offset, namely random jitter, to the human head key points, and then carrying out alignment operation on the human head image.
8. The helmet wearing detection method based on posture correction according to claim 7, characterized in that: in the step S2, the method of determining whether or not the crash helmet is worn using the output attribute of the head includes the steps of:
a1: presetting a threshold;
a2: and when the wearing confidence of the safety helmet of the head image is larger than the threshold value, judging that the person wears the safety helmet, otherwise, judging that the person does not wear the safety helmet.
CN202211356734.4A 2022-11-01 2022-11-01 Helmet wearing detection method based on attitude correction Pending CN115393905A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211356734.4A CN115393905A (en) 2022-11-01 2022-11-01 Helmet wearing detection method based on attitude correction

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211356734.4A CN115393905A (en) 2022-11-01 2022-11-01 Helmet wearing detection method based on attitude correction

Publications (1)

Publication Number Publication Date
CN115393905A true CN115393905A (en) 2022-11-25

Family

ID=84114918

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211356734.4A Pending CN115393905A (en) 2022-11-01 2022-11-01 Helmet wearing detection method based on attitude correction

Country Status (1)

Country Link
CN (1) CN115393905A (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110263686A (en) * 2019-06-06 2019-09-20 温州大学 A kind of construction site safety of image cap detection method based on deep learning
CN113052140A (en) * 2021-04-25 2021-06-29 合肥中科类脑智能技术有限公司 Video-based substation personnel and vehicle violation detection method and system
CN113158851A (en) * 2021-04-07 2021-07-23 浙江大华技术股份有限公司 Wearing safety helmet detection method and device and computer storage medium
CN113807240A (en) * 2021-09-15 2021-12-17 国网河北省电力有限公司衡水供电分公司 Intelligent transformer substation personnel dressing monitoring method based on uncooperative face recognition
CN114495191A (en) * 2021-11-30 2022-05-13 珠海亿智电子科技有限公司 Combined safety helmet wearing real-time detection method based on end side
CN115035088A (en) * 2022-06-28 2022-09-09 中国石油大学(华东) Helmet wearing detection method based on yolov5 and posture estimation

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110263686A (en) * 2019-06-06 2019-09-20 温州大学 A kind of construction site safety of image cap detection method based on deep learning
CN113158851A (en) * 2021-04-07 2021-07-23 浙江大华技术股份有限公司 Wearing safety helmet detection method and device and computer storage medium
CN113052140A (en) * 2021-04-25 2021-06-29 合肥中科类脑智能技术有限公司 Video-based substation personnel and vehicle violation detection method and system
CN113807240A (en) * 2021-09-15 2021-12-17 国网河北省电力有限公司衡水供电分公司 Intelligent transformer substation personnel dressing monitoring method based on uncooperative face recognition
CN114495191A (en) * 2021-11-30 2022-05-13 珠海亿智电子科技有限公司 Combined safety helmet wearing real-time detection method based on end side
CN115035088A (en) * 2022-06-28 2022-09-09 中国石油大学(华东) Helmet wearing detection method based on yolov5 and posture estimation

Similar Documents

Publication Publication Date Title
CN111488804B (en) Labor insurance product wearing condition detection and identity identification method based on deep learning
EP3539054B1 (en) Neural network image processing apparatus
CN108053427B (en) Improved multi-target tracking method, system and device based on KCF and Kalman
CN111898514B (en) Multi-target visual supervision method based on target detection and action recognition
CN108564069B (en) Video detection method for industrial safety helmet
CN104616438B (en) A kind of motion detection method of yawning for fatigue driving detection
CN110419048B (en) System for identifying defined objects
EP2118864B1 (en) Behavioral recognition system
JP5594791B2 (en) Attribute determination method, attribute determination apparatus, program, recording medium, and attribute determination system
JP6398979B2 (en) Video processing apparatus, video processing method, and video processing program
CN111507248B (en) Face forehead region detection and positioning method and system based on low-resolution thermodynamic diagram
US20070242856A1 (en) Object Recognition Method and Apparatus Therefor
CN110414400B (en) Automatic detection method and system for wearing of safety helmet on construction site
CN104361332A (en) Human face eye region positioning method for fatigue driving detection
CN110991315A (en) Method for detecting wearing state of safety helmet in real time based on deep learning
JP2014093023A (en) Object detection device, object detection method and program
CN110110755A (en) Based on the pedestrian of PTGAN Regional disparity and multiple branches weight recognition detection algorithm and device
CN112364778A (en) Power plant safety behavior information automatic detection method based on deep learning
Yang et al. Safety helmet wearing detection based on an improved YOLOv3 scheme
CN110688969A (en) Video frame human behavior identification method
CN114037873A (en) Transformer substation secondary circuit protection pressing plate detection system and method based on artificial intelligence
CN115393905A (en) Helmet wearing detection method based on attitude correction
CN112241695A (en) Method for recognizing portrait without safety helmet and with face recognition function
CN116385962A (en) Personnel monitoring system in corridor based on machine vision and method thereof
US20210166012A1 (en) Information processing apparatus, control method, and non-transitory storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination