CN112364734A - Abnormal dressing detection method based on yolov4 and CenterNet - Google Patents

Abnormal dressing detection method based on yolov4 and CenterNet Download PDF

Info

Publication number
CN112364734A
CN112364734A CN202011188613.4A CN202011188613A CN112364734A CN 112364734 A CN112364734 A CN 112364734A CN 202011188613 A CN202011188613 A CN 202011188613A CN 112364734 A CN112364734 A CN 112364734A
Authority
CN
China
Prior art keywords
prediction
yolov4
data
topk
centernet
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.)
Granted
Application number
CN202011188613.4A
Other languages
Chinese (zh)
Other versions
CN112364734B (en
Inventor
柯逍
陈文垚
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.)
Fuzhou University
Original Assignee
Fuzhou University
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 Fuzhou University filed Critical Fuzhou University
Priority to CN202011188613.4A priority Critical patent/CN112364734B/en
Publication of CN112364734A publication Critical patent/CN112364734A/en
Application granted granted Critical
Publication of CN112364734B publication Critical patent/CN112364734B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Multimedia (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Human Computer Interaction (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to an abnormal dressing detection method based on yolov4 and CenterNet, which comprises the following steps: s1, acquiring abnormal dressing data and preprocessing the abnormal dressing data to construct an abnormal dressing detection data set; step S2, respectively optimizing and respectively training the hyperparameters of the yolov4 model and the CenterNet model; s3, carrying out target detection according to the trained yolov4 detection model to obtain a prediction result, decoding the prediction result, and screening out a final prediction frame by utilizing non-maximum inhibition; step S4, carrying out target detection according to the trained CenterNet detection model, calculating the center point and corner point thermodynamic diagrams of input data, predicting the positions of key points, normalizing the obtained thermodynamic diagrams, and finding K maximum value points as candidate targets; calculating through the central point to obtain a final target frame; and step S5, drawing the final prediction frame and the final target frame on the original image, and adding information such as category names, confidence degrees and the like to obtain a result image. The invention can effectively identify the abnormal dressing phenomenon.

Description

Abnormal dressing detection method based on yolov4 and CenterNet
Technical Field
The invention relates to the field of image recognition and computer vision, in particular to an abnormal dressing detection method based on yolov4 and CenterNet.
Background
In the process of daily production and life, certain specific occasions have strict requirements on dressing of people, so that the safety of people in the process of production and life is ensured. In a construction site, a constructor needs to wear protective devices such as safety helmets, safety vests and the like to ensure the working safety of the constructor. The construction industry is one of the dangerous industries in all industries, and data of the American labor statistical service (BLS) shows that the number of deaths in the construction industry is far higher than that in all other industries, and the fatal injury rate is also higher than the national average level of all the industries. The responsible personnel can not supervise the force, the constructors can not strictly abide by the relevant safety regulations of the construction site, and safety protection equipment such as safety helmets, safety vests and the like are worn, so that the safety accidents are frequently caused. In hospitals, medical staff and various infectious patients are very easy to contact, and have higher infection risk. Respiratory infectious diseases are one of the most dangerous diseases in various infectious diseases at present, and the special transmission path of the respiratory infectious diseases causes the respiratory infectious diseases to be difficult to prevent and control. Medical personnel and patients can be protected from infection by a simple but very useful method of wearing a mask which provides a bi-directional protective effect between the medical personnel and the patient. The wearing of the mask is of great significance for protecting medical care personnel and patients and preventing infectious diseases from spreading widely. In these special occasions, abnormal dressing behaviors such as wearing no safety helmet, safety vest, or gauze mask can greatly increase the danger in the production and life processes of people.
At present, the identification of the abnormal dressing condition is mainly based on a manual inspection mode. The manual inspection is easily interfered by various factors to reduce the efficiency, and the all-weather monitoring can not be carried out and the human resources are wasted, so the efficiency of the method is low, and the actual requirements of all safety supervision departments can not be fully met. To monitor the abnormal dressing phenomenon in real time, the abnormal dressing phenomenon can be automatically identified by monitoring equipment such as a camera and a detection algorithm based on machine vision, so that the abnormal dressing condition can be monitored in real time, the problems of missed detection, false detection and the like caused by various subjective factors in manual detection are avoided, and the efficiency and the automation level of supervision can be improved while manpower resources are saved.
Disclosure of Invention
In view of the above, the present invention provides an abnormal dressing detection method based on yolov4 and centrnet, which can effectively identify the abnormal dressing phenomenon, and can maintain a relatively good detection speed while achieving a good detection accuracy.
In order to achieve the purpose, the invention adopts the following technical scheme:
an abnormal dressing detection method based on yolov4 and CenterNet comprises the following steps:
s1, acquiring abnormal dressing data, preprocessing the abnormal dressing data, and respectively constructing an abnormal dressing detection data set according to yolov4 model requirements and CenterNet model requirements;
step S2, respectively optimizing the hyperparameters of the yolov4 model and the CenterNet model, and respectively training according to corresponding data sets to obtain a yolov4 detection model and a CenterNet detection model;
s3, carrying out target detection according to the trained yolov4 detection model to obtain a prediction result, decoding the prediction result, and screening out a final prediction frame by utilizing non-maximum inhibition;
step S4, carrying out target detection according to the trained CenterNet detection model, calculating the center point and corner point thermodynamic diagrams of input data, predicting the positions of key points, normalizing the obtained thermodynamic diagrams, and finding K maximum value points as candidate targets; calculating through the central point to obtain a final prediction frame;
and step S5, drawing the final prediction frames obtained respectively according to the two detection models on the original image, and adding information such as category names, confidence degrees and the like to obtain a result image.
Further, in step S1, specifically, the step includes:
step S11, acquiring data pictures related to abnormal dresses, screening data and finishing data preprocessing;
step S12: manufacturing a data set according to the requirements of a yolov4 model, dividing all data into a training set and a testing set according to a proportion, and generating a txt file required by a training model according to an xml file containing image data marking information;
step S13: and (3) making a data set according to the requirements of the CenterNet model, dividing all data into a training set, a testing set and a verification set according to a proportion, and generating a json file required by the training model according to the xml file containing the image data labeling information.
Further, in step S2, specifically, the step includes:
step S21: acquiring an optimal value of the super-parameter, and adjusting the super-parameter to make the training model reach the best;
step S22: setting the momentum parameter momentum in the momentum gradient descent to 0.9 and setting the regularization coefficient of weight attenuation to 0.0005 in the training file of yolov 4;
step S23: a step method is adopted for the learning rate adjustment mode;
step S24: calculating an anchor by using a k-means clustering algorithm, and normalizing the width and the height of the bounding box by using the width and the height of the data picture, namely:
Figure BDA0002752083160000031
let anchor ═ wa,ha),box＝(wb,hb) Using the IOU as a metric, the calculation is as follows:
Figure BDA0002752083160000032
the value of the IOU is between 0 and 1, the more similar the two box are, the larger the IOU value is, and the final measurement formula is as follows:
d(box,anchor)＝1-IOU(box,anchor)
randomly selecting k bounding boxes in the data set as initial anchors, using IOU measurement to allocate each bounding box to the anchor closest to the bounding box, traversing all the bounding boxes, calculating the average value of the width and the height of all the bounding boxes in each cluster, updating the anchors, and repeating the steps until the anchors are not changed or the maximum iteration number is reached.
Further, in step S3, specifically, the step includes:
step S31: after input data passes through a feature extraction network darknet53, three feature maps with different sizes are obtained;
step S32, according to the three extracted feature maps with different sizes, one part of the results obtained by convolving the three initial feature maps is used for outputting the prediction results corresponding to the feature maps, and the other part of the results is used for combining with other feature maps after deconvolution, so that the prediction results of the three effective feature maps are finally obtained;
step S33: adjusting a preset prior frame according to the obtained prediction result to obtain the size and the position information of the prediction frame;
step S34: according to the adjusted prediction frame, carrying out maximum suppression, carrying out local search in the candidate target to find the prediction frame with the highest confidence coefficient and suppress the prediction frame with the lower confidence coefficient,
further, in step S33, specifically, the step includes:
(a) the characteristic diagram is divided into S multiplied by S grids, and then a preset prior frame is adjusted to the effective characteristic diagram
(b) Then obtaining coordinate information x _ offset, y _ offset, h and w of a prior frame from the yolov4 network prediction result;
(c) and (3) carrying out sigmoid function processing on the center point coordinates of the prior frame corresponding to the grid, adding corresponding x _ offset and y _ offset to obtain the center of the prediction frame, calculating by using h and w to obtain the width and height of the prediction frame, and finally obtaining the size and position information of the prediction frame.
Further, in step S34, specifically, the step includes:
(a) when non-maximum value suppression is carried out, the prediction frames of the same target are sorted from high confidence to low confidence, the prediction frame with the highest confidence is taken out to be respectively used for calculating the IOU with the other prediction frames,
setting two detection boxes B according to the process of finding local maximum value by using intersection ratio IOU1And B2Then the intersection between the two is as follows:
Figure BDA0002752083160000051
(b) and if the calculation result is greater than the set threshold value, the prediction frame is restrained and cannot be output as a result, and after all the prediction frames are calculated, the prediction frame with the maximum confidence coefficient in the rest prediction frames is taken out to repeat the operation.
Further, the step S4 specifically includes the following steps:
step S41: extracting the maximum values of the central point in the horizontal direction and the vertical direction through centerpooling, and adding the maximum values to provide information except the position of the central point, so as to obtain a central point thermodynamic diagram;
step S42: predicting the upper left corner point and the lower right corner point of the object by using a cascade coder firing and extracting the maximum value max of the boundary of the object1Finding the maximum max inwards along the boundary maximum2Then, adding and capturing the boundary features to provide semantic information of the associated object for the corner features so as to obtain a corner thermodynamic diagram;
step S43: normalizing the obtained thermodynamic diagram by using a sigmoid function, wherein the formula of the sigmoid function is as follows:
Figure BDA0002752083160000052
step S44: performing 3 × 3 maximum pooling convolution on the thermodynamic diagram, finding and screening the top k maximum points by a _ topk function to represent the existence target, and obtaining topk _ score, topk _ inds, topk _ clses, topk _ ys and topk _ xs, wherein topk _ score represents the top k maximum scores of each category, topk _ inds represents the lower corner marks of k scores corresponding to each category, top k maximum scores of all categories of topk _ clses, and topk _ ys and topk _ xs are used for extracting results from corresponding tensors according to the lower corner marks;
step S45: and calculating the box and the corresponding category and confidence coefficient thereof according to the topk _ ys and the topk _ xs.
Further, the step S5 is specifically:
step S51: calculating to obtain the coordinates of the upper left corner point of the prediction frame according to the coordinates of the center point and the width and the height of the center point so as to obtain the position information of the prediction frame in the output picture;
step S52: and drawing the prediction frame, the prediction category and the confidence coefficient on an output picture by using a drawing function to obtain a final result.
Compared with the prior art, the invention has the following beneficial effects:
the invention can effectively identify the abnormal dressing phenomenon, and can keep relatively good detection speed while achieving good detection precision.
Drawings
Fig. 1 is a schematic diagram of the principle of the present invention.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
Referring to fig. 1, the present invention provides an abnormal wearing detection method based on yolov4 and centrnet, comprising the following steps:
s1, acquiring abnormal dressing data, preprocessing the abnormal dressing data, and respectively constructing an abnormal dressing detection data set according to yolov4 model requirements and CenterNet model requirements;
step S2, respectively optimizing the hyperparameters of the yolov4 model and the CenterNet model, and respectively training according to corresponding data sets to obtain a yolov4 detection model and a CenterNet detection model;
step S3, carrying out target detection according to the trained yolov4 detection model, extracting a feature map by a trunk feature extraction network darknet53, carrying out operations such as convolution and the like on the feature map, predicting the feature map to obtain a prediction result, and screening out a final prediction frame by utilizing non-maximum value inhibition after decoding the prediction result;
step S4, carrying out target detection according to the trained CenterNet detection model, calculating the center point and corner point thermodynamic diagrams of input data, predicting the positions of key points, normalizing the obtained thermodynamic diagrams, and finding K maximum value points as candidate targets; calculating through the central point to obtain a final prediction frame;
and step S5, drawing the final prediction frames obtained respectively according to the two detection models on the original image, and adding information such as category names, confidence degrees and the like to obtain a result image.
In this embodiment, the step S1 specifically includes:
step S11, acquiring data pictures related to abnormal dresses, screening data and finishing data preprocessing;
step S12: manufacturing a data set according to the requirements of a yolov4 model, dividing all data into a training set and a testing set according to a proportion, and generating a txt file required by a training model according to an xml file containing image data marking information; the txt file comprises the category and position information of each target in the picture data;
step S13: and manufacturing a data set according to the requirements of the CenterNet model, dividing all data into a training set, a testing set and a verification set according to a proportion, and generating a json file required by the training model according to the xml file containing the image data labeling information, wherein the json file contains information such as an image name, an image size, a target category, a position coordinate, a size and the like in the image data.
In this embodiment, the step S2 specifically includes:
step S21: obtaining the optimal value of the hyper-parameters in the relevant training files according to multiple experiments, and optimizing the hyper-parameters to enable the training model to be optimal;
step S22: setting the momentum parameter momentum in the momentum gradient descent to 0.9 in the training file of yolov4 can effectively prevent the loss function from falling into the local minimum value in the network training, accelerate the speed of gradient convergence to the optimal value, and set the regular coefficient decay of weight attenuation to 0.0005, thereby effectively preventing overfitting;
step S23: if the learning rate is too large, the weight updating speed is high, but the selection of the optimal value is easy to miss, if the learning rate is too small, the weight updating speed is slow, the training efficiency is low, the training speed and the selection of the optimal value can be effectively improved by setting a suitable learning rate, the learning rate is adjusted by adopting a steps method, and the learning rate is attenuated by a certain multiple when a certain number of iterations is reached;
step S24: calculating an anchor by using a k-means clustering algorithm, and normalizing the width and the height of the bounding box by using the width and the height of the data picture, namely:
Figure BDA0002752083160000081
let anchor ═ wa,ha),box＝(wb,hb) Using the IOU as a metric, the calculation is as follows:
Figure BDA0002752083160000082
the value of the IOU is between 0 and 1, the more similar the two box are, the larger the IOU value is, and the final measurement formula is as follows:
d(box,anchor)＝1-IOU(box,anchor)
randomly selecting k bounding boxes in the data set as initial anchors, using IOU measurement to allocate each bounding box to the anchor closest to the bounding box, traversing all the bounding boxes, calculating the average value of the width and the height of all the bounding boxes in each cluster, updating the anchors, and repeating the steps until the anchors are not changed or the maximum iteration number is reached.
In this embodiment, the step S3 specifically includes:
step S31: the darknet53 is a backbone of yolov4, after data is input, the darknet53 firstly uses Conv2D to process the input data, L2 regularization is carried out during convolution, batch standardization is carried out after the convolution is finished, a LeakyReLU activation function is used, then, a residual convolution block in the darknet53 is used for carrying out convolution on the output once by 3 x 3 and with the step length of 2, then, the convolution layer is stored and is carried out again by carrying out convolution by 1 x 1 and 3 x 3, and the result and the stored convolution layer are added to serve as a final result, so that the depth of the neural network is increased, and the problem of gradient disappearance caused by the increase of the depth in the deep neural network is relieved. After the input data passes through the feature extraction network darknet53, three feature maps with different sizes are obtained, namely (52, 256), (26, 512), (13, 1024);
step S32 is to extract three feature maps of different sizes through the feature extraction network darknet53, and a part of results obtained by performing operations such as convolution on the three initial feature maps is used to output a prediction result corresponding to the feature map, and another part is used to perform deconvolution and then combine with other feature maps. Finally, the prediction results of three effective feature maps are obtained, the sizes of the prediction results are (255,13,13), (255,26,13) and (255,52,52), and the positions of 3 prediction frames on the grids which are divided into 13 × 13, 26 × 26 and 52 × 52 corresponding to each map are obtained;
step S33: the prediction result obtained after processing the feature map is not a final prediction frame, a preset prior frame needs to be adjusted by using the prediction result of the yolov4 network, the feature map is firstly divided into S multiplied by S grids, the preset prior frame is adjusted to an effective feature map, then coordinate information x _ offset, y _ offset, h and w of the prior frame is obtained from the prediction result of the yolov4 network, the center point coordinate of the prior frame corresponding to the grid is subjected to sigmoid function processing and then is added with the corresponding x _ offset and y _ offset to obtain the center of the prediction frame, then the width and height of the prediction frame are obtained by using h and w to calculate, and finally the size and position information of the prediction frame are obtained;
step S34: each target usually has a plurality of prediction boxes, and maximum suppression is required to be performed to search for a prediction box with the highest confidence degree and suppress a prediction box with a lower confidence degree in a local search among candidate targets, which is essentially a process of searching for a local maximum value according to an intersection ratio IOU. The cross-over ratio IOU is an important index for measuring the overlapping degree of two prediction frames, and is provided with two detection frames B1And B2Then the intersection between the two is as follows:
Figure BDA0002752083160000101
when the non-maximum value is restrained, firstly, the prediction frames of the same target are sorted from large to small according to the confidence level, then the prediction frame with the highest confidence level is taken out to be respectively compared with the rest prediction frames to calculate the IOU, if the calculation result is larger than the set threshold value, the prediction frame is restrained and cannot be output as a result, and after all the prediction frames are calculated, the prediction frame with the highest confidence level in the rest prediction frames is taken out to repeat the operation.
In this embodiment, the step S4 specifically includes the following steps:
step S41: extracting the maximum values of the central point in the horizontal direction and the vertical direction through centerpooling, and adding the maximum values to provide information except the position of the central point, so as to obtain a central point thermodynamic diagram;
step S42: cascade corerpoling is utilized to predict upper left corner and lower right corner points of the object and extract the maximum value max of the object boundary1Finding the maximum max inwards along the boundary maximum2Then, adding and capturing the boundary features to provide semantic information of the associated object for the corner features so as to obtain a corner thermodynamic diagram;
step S43: normalizing the obtained thermodynamic diagram by using a sigmoid function, wherein the formula of the sigmoid function is as follows:
Figure BDA0002752083160000102
step S44: performing 3 × 3 maximum pooling convolution on the thermodynamic diagram, finding and screening the top k maximum points by a _ topk function to represent the existence target, and obtaining topk _ score, topk _ inds, topk _ clses, topk _ ys and topk _ xs, wherein topk _ score represents the top k maximum scores of each category, topk _ inds represents the lower corner marks of k scores corresponding to each category, top k maximum scores of all categories of topk _ clses, and topk _ ys and topk _ xs are used for extracting results from corresponding tensors according to the lower corner marks;
step S45: and calculating the box and the corresponding category and confidence coefficient thereof according to the topk _ ys and the topk _ xs.
In this embodiment, the step S5 specifically includes:
step S51: calculating to obtain the coordinates of the upper left corner point of the prediction frame according to the coordinates of the center point and the width and the height of the center point so as to obtain the position information of the prediction frame in the output picture;
step S52: and drawing the prediction frame, the prediction category and the confidence coefficient on an output picture by using a drawing function to obtain a final result.
The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.

Claims (8)

1. An abnormal dressing detection method based on yolov4 and CenterNet is characterized by comprising the following steps:
s1, acquiring abnormal dressing data, preprocessing the abnormal dressing data, and respectively constructing an abnormal dressing detection data set according to yolov4 model requirements and CenterNet model requirements;
step S2, respectively optimizing the hyperparameters of the yolov4 model and the CenterNet model, and respectively training according to corresponding data sets to obtain a yolov4 detection model and a CenterNet detection model;
s3, carrying out target detection according to the trained yolov4 detection model to obtain a prediction result, decoding the prediction result, and screening out a final prediction frame by utilizing non-maximum inhibition;
step S4, carrying out target detection according to the trained CenterNet detection model, calculating the center point and corner point thermodynamic diagrams of input data, predicting the positions of key points, normalizing the obtained thermodynamic diagrams, and finding K maximum value points as candidate targets; calculating through the central point to obtain a final prediction frame;
and step S5, drawing the final prediction frames obtained respectively according to the two detection models on the original image, and adding information such as category names, confidence degrees and the like to obtain a result image.
2. The abnormal dressing detection method based on yolov4 and centrnet of claim 1, wherein the step S1 specifically comprises:
step S11, acquiring data pictures related to abnormal dresses, screening data and finishing data preprocessing;
step S12: manufacturing a data set according to the requirements of a yolov4 model, dividing all data into a training set and a testing set according to a proportion, and generating a txt file required by a training model according to an xml file containing image data labeling information;
step S13: and (3) making a data set according to the requirements of the CenterNet model, dividing all data into a training set, a testing set and a verification set according to a proportion, and generating a json file required by the training model according to the xml file containing the image data labeling information.
3. The abnormal dressing detection method based on yolov4 and centrnet of claim 1, wherein the step S2 specifically comprises:
step S21: acquiring an optimal value of the hyper-parameter, and adjusting the hyper-parameter to optimize the training model;
step S22: setting the momentum parameter momentum in the momentum gradient descent to 0.9 and setting the regularization coefficient of weight attenuation to 0.0005 in the training file of yolov 4;
step S23: a step method is adopted for the learning rate adjustment mode;
step S24: calculating an anchor by using a k-means clustering algorithm, and normalizing the width and the height of the bounding box by using the width and the height of the data picture, namely:
Figure FDA0002752083150000021
let anchor ═ wa,ha),box＝(wb,hb) Using the IOU as a metric, the calculation is as follows:
Figure FDA0002752083150000022
the value of the IOU is between 0 and 1, the more similar the two box are, the larger the value of the IOU is, and the final measurement formula is as follows:
d(box,anchor)＝1-IOU(box,anchor)
randomly selecting k bounding boxes in the data set as initial anchors, using IOU measurement to allocate each bounding box to the anchor closest to the bounding box, traversing all the bounding boxes, calculating the average value of the width and the height of all the bounding boxes in each cluster, updating the anchors, and repeating the steps until the anchors are not changed or the maximum iteration number is reached.
4. The abnormal dressing detection method based on yolov4 and centrnet of claim 1, wherein the step S3 specifically comprises:
step S31: after input data passes through a feature extraction network darknet53, three feature maps with different sizes are obtained;
step S32, according to the extracted feature maps with different sizes, one part of the results obtained by convolving the three initial feature maps is used for outputting the prediction results corresponding to the feature maps, and the other part of the results is used for combining with other feature maps after deconvolution, so that the prediction results of the three effective feature maps are finally obtained;
step S33: adjusting a preset prior frame according to the obtained prediction result to obtain the size and the position information of the prediction frame;
step S34: and according to the adjusted prediction frame, carrying out maximum suppression, carrying out local search in the candidate target to find the prediction frame with the highest confidence coefficient and suppressing the prediction frame with the lower confidence coefficient.
5. The abnormal dressing detection method based on yolov4 and CenterNet of claim 4, wherein the step S33 specifically comprises:
(a) the characteristic diagram is divided into S multiplied by S grids, and then a preset prior frame is adjusted to the effective characteristic diagram
(b) Then obtaining coordinate information x _ offset, y _ offset, h and w of a prior frame from the yolov4 network prediction result;
(c) and (3) carrying out sigmoid function processing on the center point coordinates of the prior frame corresponding to the grid, adding corresponding x _ offset and y _ offset to obtain the center of the prediction frame, calculating by using h and w to obtain the width and height of the prediction frame, and finally obtaining the size and position information of the prediction frame.
6. The abnormal dressing detection method based on yolov4 and CenterNet of claim 4, wherein the step S34 specifically comprises:
(a) when non-maximum value suppression is carried out, the prediction frames of the same target are sorted from high confidence to low confidence, the prediction frame with the highest confidence is taken out to be respectively used for calculating the IOU with the other prediction frames,
setting two detection boxes B according to the process of finding local maximum value by using intersection ratio IOU1And B2Then the intersection between the two is as follows:
Figure FDA0002752083150000041
(b) and if the calculation result is greater than the set threshold value, the prediction frame is restrained and cannot be output as a result, and after all the prediction frames are calculated, the prediction frame with the maximum confidence coefficient in the rest prediction frames is taken out to repeat the operation.
7. The abnormal dressing detection method based on yolov4 and CenterNet as claimed in claim 1, wherein said step S4 specifically comprises the following steps:
step S41: extracting the maximum values of the central point in the horizontal direction and the vertical direction through center firing, and adding the maximum values to provide information except the position of the central point, thereby obtaining a central point thermodynamic diagram;
step S42: cascade coder firing is utilized to predict the upper left corner point and the lower right corner point of the object and extract the maximum value max of the boundary of the object1Finding the maximum max inwards along the boundary maximum2Then, adding and capturing the boundary features to provide semantic information of the associated object for the corner features so as to obtain a corner thermodynamic diagram;
step S43: normalizing the obtained thermodynamic diagram by using a sigmoid function, wherein the formula of the sigmoid function is as follows:
Figure FDA0002752083150000042
step S44: performing 3 × 3 maximum pooling convolution on the thermodynamic diagram, finding and screening the top k maximum points by a _ topk function to represent the existence target, and obtaining topk _ score, topk _ inds, topk _ clses, topk _ ys and topk _ xs, wherein topk _ score represents the top k maximum scores of each category, topk _ inds represents the lower corner marks of k scores corresponding to each category, the top k maximum scores of all categories of topk _ clses, and topk _ ys and topk _ xs are used for extracting results from corresponding tensors according to the lower corner marks;
step S45: and calculating the box and the corresponding category and confidence coefficient thereof according to the topk _ ys and the topk _ xs.
8. The abnormal dressing detection method based on yolov4 and CenterNet as claimed in claim 1, wherein the step S5 specifically comprises:
step S51: calculating to obtain the coordinates of the upper left corner point of the prediction frame according to the coordinates of the center point and the width and the height of the center point so as to obtain the position information of the prediction frame in the output picture;
step S52: and drawing the prediction frame, the prediction category and the confidence coefficient on an output picture by using a drawing function to obtain a final result.
CN202011188613.4A 2020-10-30 2020-10-30 Abnormal dressing detection method based on yolov4 and CenterNet Active CN112364734B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011188613.4A CN112364734B (en) 2020-10-30 2020-10-30 Abnormal dressing detection method based on yolov4 and CenterNet

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011188613.4A CN112364734B (en) 2020-10-30 2020-10-30 Abnormal dressing detection method based on yolov4 and CenterNet

Publications (2)

Publication Number Publication Date
CN112364734A true CN112364734A (en) 2021-02-12
CN112364734B CN112364734B (en) 2023-02-21

Family

ID=74513822

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011188613.4A Active CN112364734B (en) 2020-10-30 2020-10-30 Abnormal dressing detection method based on yolov4 and CenterNet

Country Status (1)

Country Link
CN (1) CN112364734B (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112915539A (en) * 2021-04-01 2021-06-08 腾讯科技（深圳）有限公司 Virtual object detection method and device and readable storage medium
CN112990131A (en) * 2021-04-27 2021-06-18 广东科凯达智能机器人有限公司 Method, device, equipment and medium for acquiring working gear of voltage change-over switch
CN113111754A (en) * 2021-04-02 2021-07-13 中国科学院深圳先进技术研究院 Target detection method, device, terminal equipment and storage medium
CN113408624A (en) * 2021-06-22 2021-09-17 福州大学 Task-oriented image quality evaluation method based on transfer learning
CN113837138A (en) * 2021-09-30 2021-12-24 重庆紫光华山智安科技有限公司 Dressing monitoring method, system, medium and electronic terminal
CN114882556A (en) * 2022-04-26 2022-08-09 西北大学 Method for detecting makeup face of opera character based on improved YooloX

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110532894A (en) * 2019-08-05 2019-12-03 西安电子科技大学 Remote sensing target detection method based on boundary constraint CenterNet
US20200250487A1 (en) * 2017-10-23 2020-08-06 Hangzhou Hikvision Digital Technology Co., Ltd. Target detection method and apparatus, and computer device
CN111553348A (en) * 2020-04-26 2020-08-18 中南大学 Anchor-based target detection method based on centernet
CN111626277A (en) * 2020-08-03 2020-09-04 杭州智诚惠通科技有限公司 Vehicle tracking method and device based on over-station inter-modulation index analysis
CN111709414A (en) * 2020-06-29 2020-09-25 济南浪潮高新科技投资发展有限公司 AR device, character recognition method and device thereof, and computer-readable storage medium
CN111709489A (en) * 2020-06-24 2020-09-25 广西师范大学 Citrus identification method based on improved YOLOv4
CN111738206A (en) * 2020-07-08 2020-10-02 浙江浙能天然气运行有限公司 Excavator detection method for unmanned aerial vehicle inspection based on CenterNet
CN111814742A (en) * 2020-07-29 2020-10-23 南方电网数字电网研究院有限公司 Knife switch state identification method based on deep learning
CN111832513A (en) * 2020-07-21 2020-10-27 西安电子科技大学 Real-time football target detection method based on neural network

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200250487A1 (en) * 2017-10-23 2020-08-06 Hangzhou Hikvision Digital Technology Co., Ltd. Target detection method and apparatus, and computer device
CN110532894A (en) * 2019-08-05 2019-12-03 西安电子科技大学 Remote sensing target detection method based on boundary constraint CenterNet
CN111553348A (en) * 2020-04-26 2020-08-18 中南大学 Anchor-based target detection method based on centernet
CN111709489A (en) * 2020-06-24 2020-09-25 广西师范大学 Citrus identification method based on improved YOLOv4
CN111709414A (en) * 2020-06-29 2020-09-25 济南浪潮高新科技投资发展有限公司 AR device, character recognition method and device thereof, and computer-readable storage medium
CN111738206A (en) * 2020-07-08 2020-10-02 浙江浙能天然气运行有限公司 Excavator detection method for unmanned aerial vehicle inspection based on CenterNet
CN111832513A (en) * 2020-07-21 2020-10-27 西安电子科技大学 Real-time football target detection method based on neural network
CN111814742A (en) * 2020-07-29 2020-10-23 南方电网数字电网研究院有限公司 Knife switch state identification method based on deep learning
CN111626277A (en) * 2020-08-03 2020-09-04 杭州智诚惠通科技有限公司 Vehicle tracking method and device based on over-station inter-modulation index analysis

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
谈小峰等: "基于YOLOv4改进算法的乒乓球识别", 《科技创新与应用》, no. 27, 8 September 2020 (2020-09-08) *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112915539A (en) * 2021-04-01 2021-06-08 腾讯科技（深圳）有限公司 Virtual object detection method and device and readable storage medium
CN112915539B (en) * 2021-04-01 2023-01-06 腾讯科技（深圳）有限公司 Virtual object detection method and device and readable storage medium
CN113111754A (en) * 2021-04-02 2021-07-13 中国科学院深圳先进技术研究院 Target detection method, device, terminal equipment and storage medium
CN112990131A (en) * 2021-04-27 2021-06-18 广东科凯达智能机器人有限公司 Method, device, equipment and medium for acquiring working gear of voltage change-over switch
CN113408624A (en) * 2021-06-22 2021-09-17 福州大学 Task-oriented image quality evaluation method based on transfer learning
CN113837138A (en) * 2021-09-30 2021-12-24 重庆紫光华山智安科技有限公司 Dressing monitoring method, system, medium and electronic terminal
CN113837138B (en) * 2021-09-30 2023-08-29 重庆紫光华山智安科技有限公司 Dressing monitoring method, dressing monitoring system, dressing monitoring medium and electronic terminal
CN114882556A (en) * 2022-04-26 2022-08-09 西北大学 Method for detecting makeup face of opera character based on improved YooloX
CN114882556B (en) * 2022-04-26 2024-03-15 西北大学 Improved YoloX-based drama character dressing face detection method

Also Published As

Publication number Publication date
CN112364734B (en) 2023-02-21

Similar Documents

Publication Publication Date Title
CN112364734B (en) Abnormal dressing detection method based on yolov4 and CenterNet
CN110399905B (en) Method for detecting and describing wearing condition of safety helmet in construction scene
CN103955699B (en) A kind of real-time fall events detection method based on monitor video
CN110969244B (en) Building construction safety monitoring method based on convolutional neural network
CN108960056A (en) A kind of fall detection method based on posture analysis and Support Vector data description
CN110070010A (en) A kind of face character correlating method identified again based on pedestrian
CN106997629A (en) Access control method, apparatus and system
US20220232247A1 (en) Image coding method, action recognition method, and action recognition apparatus
CN111062303A (en) Image processing method, system and computer storage medium
CN110414400A (en) A kind of construction site safety cap wearing automatic testing method and system
CN112580995A (en) Construction safety big data monitoring system and safety risk dynamic evaluation method
CN109948501A (en) The detection method of personnel and safety cap in a kind of monitor video
CN110287825A (en) It is a kind of that motion detection method is fallen down based on crucial skeleton point trajectory analysis
CN115294533B (en) Building construction state monitoring method based on data processing
Rentao et al. Indoor smoking behavior detection based on yolov3-tiny
CN109308444A (en) A kind of abnormal behaviour recognition methods under indoor environment
CN114973320A (en) Underground coal mine personnel detection method based on depth information
CN117475502B (en) Iris and face fusion recognition method and system based on mine
CN114360209A (en) Video behavior identification security system based on deep learning
Yi et al. Research on Helmet Wearing Detection in Multiple Scenarios Based on YOLOv5
CN114359831A (en) Risk omen reasoning-oriented intelligent identification system and method for worker side-falling
CN113111808B (en) Abnormal behavior detection method and system based on machine vision
CN115469626A (en) Industrial production monitoring system and method based on Internet of things
KR101781359B1 (en) A Method Of Providing For Searching Footprint And The System Practiced The Method
CN115170998A (en) Human behavior and action recognition method and device, computer equipment and 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
GR01 Patent grant
GR01 Patent grant