KR102136070B1 - Disaster and Accident Prevention Device in Construction site using AI image analysis - Google Patents

Abstract

The present invention relates to an apparatus for preventing a disaster and an accident in a construction site by using artificial intelligence image analysis, capable of enabling independent installation in a dangerous zone of a construction site and making a warning with proper content by analyzing an image of a worker approaching the dangerous zone. According to the present invention, the apparatus for preventing a disaster and an accident in a construction site by using artificial intelligence image analysis includes: a housing having an internal storage space; a camera installed outside the housing and photographing an image from a predetermined angle range; a control part provided in the storage space and analyzing the image photographed through the camera in real time; and a speaker selectively outputting a warning announcement due to the control of the control part. The control part recognizes dangerous behavior from behavior patterns of a worker by repetitively learning the image photographed through the camera and determining whether safety equipment is used through object recognition.

Description

Disaster and Accident Prevention Device in Construction site using AI image analysis}

The present invention relates to an apparatus for preventing disasters and accidents at a construction site using artificial intelligence image analysis, and more specifically, to detect whether a worker is wearing a protective gear and whether it is installed in a dangerous element such as a dangerous object or opening in a work site. And, if the worker is within the danger range, it guides (warns) and prevents accidents and accidents at the construction site using artificial intelligence image analysis.

Among construction disasters, disasters are concentrated in small-scale construction sites, as the number of accidents (2014) at small construction sites with construction costs of less than 2 billion won account for 72% (16,489) of total construction accidents.

According to the occurrence status of each type of fatal accident by the Health and Safety Corporation, it was found that 73 deaths, 73.7% of the total in construction, and 16 deaths, 16.2% in civil engineering.

It is estimated that, due to these construction disasters, more than 20,000 injuries and more than 600 deaths occur every year at construction sites over the past decade, and national losses are expected to exceed 12 trillion won per year.

In addition, construction disasters are not only a direct economic loss, but also deteriorating national credibility and aggravating social anxiety.

On the other hand, despite numerous countermeasures, many fall accidents and safety accidents have occurred since workers are frequently using smartphones, etc., on the move or during work. Various techniques have been developed to do this.

In the prior art as described above, the registered patent publication No. 1706221 discloses a surveillance camera and an alarm control method (hereinafter referred to as'patent document') through object identification of the surveillance camera.

The alarm control method disclosed in the patent document includes the steps of: detecting, by a surveillance camera, an object having a motion from an input image; Identifying whether the detected object where the motion has occurred is an object to be monitored or disturbed; And controlling an alarm when the identified target is a surveillance target, and configured to remove the generation of an incorrect alarm through target identification of the surveillance camera.

However, the alarm control method disclosed in the patent document transmits a live view image to a video recorder through a video signal channel (SVID) while communicating an image captured by a surveillance camera with a digital video recorder (DVR) as a recording device through a communication channel. As it is constructed, it is difficult to establish its own communication network at the construction site.

That is, a surveillance camera provided at a construction site detects a target having a movement from an input image and transmits the corresponding image to a digital processing signal at a remote location through a communication network, and analyzes the image received from the digital processing signal to control the alarm. As it is configured, an additional configuration such as a communication device must be provided, whereby there is a problem in that alarm control cannot be properly performed when there is an error in the communication network.

Therefore, it is required to develop a disaster and accident prevention device that can notify a worker in a timely manner and approach a danger area without building a complex communication network.

KR 10-1906221 B1 (2017. 02. 07.) KR 10-1951119 B1 (February 15, 2019) KR 10-0805563 B1 (February 13, 2008) KR 10-1999-0047360 A (July 05, 1999)

The present invention has been devised to solve the problems of the conventional danger zone notification device as described above, and the problem to be solved in the present invention can be independently installed in the danger zone of the construction site, and the workers wear protective gear and the danger zone It is to provide a disaster prevention and accident prevention device for construction sites using artificial intelligence image analysis that can guide warnings of appropriate contents by analyzing the image of a worker approaching with.

In order to solve the above problems, a disaster and accident prevention apparatus for a construction site using artificial intelligence image analysis according to the present invention includes a housing provided with an accommodation space therein; A camera that is installed on the outside of the housing and shoots an image in a predetermined angle range; A control unit that is provided in the accommodation space and analyzes an image shot through the camera in real time; And a speaker that selectively outputs a warning broadcast under the control of the control unit, and the control unit repeatedly learns an image photographed through the camera to recognize danger behavior from an operator's behavior pattern and protects it through object recognition. It is characterized by determining whether or not the equipment is operating.

And in the present invention, the controller uses a TensorFlow library to repeatedly learn an image, a multi-layer perceptron capable Keras is applied, and a convolutional neural network (CNN) to recognize an image. Another feature is that it is used.

In addition, another aspect of the present invention is that the controller generates a data set necessary for initial learning based on a whole body image of a general person using the image data generator of Keras.

In addition, according to another aspect of the present invention, the control unit extracts a feature for an input image image using the generated dataset, and classifies according to the feature through a neural network based on the extracted feature. do.

In the present invention, the controller distinguishes an operator from an image input through the camera and then additionally determines whether or not to wear the protective equipment, and the warning broadcast is selected according to the movement path of the worker and the wearing condition of the protective equipment to select the speaker. Another feature is to be output through.

According to the present invention, an appropriate warning broadcast can be immediately output to a worker approaching the danger zone without being installed in the danger zone of the construction site without control of the safety manager at the construction site, thereby further guaranteeing the safety of the worker at the construction site. There is an advantage to do.

In addition, since the movement direction of the worker approaching the danger zone is tracked to accurately determine whether or not the user approaches the danger zone, unnecessary warning broadcasts can be prevented from being output unnecessarily.

1 is a block diagram showing an example of an apparatus for preventing disasters and accidents in a construction site using artificial intelligence image analysis according to the present invention.
Figure 2 is a view showing an example of a disaster prevention and accident prevention device installed on the wall of the construction site using artificial intelligence image analysis according to the present invention.
3 is a view showing a shooting range of the camera according to the present invention.
4 is a view showing an example of a disaster prevention and accident prevention device in a construction site using an artificial intelligence image analysis according to the present invention in a hazardous area.
5 is a flowchart showing an example of recognizing workers and selectively outputting a warning broadcast through a disaster and accident prevention device at a construction site using artificial intelligence image analysis according to the present invention.

Hereinafter will be described in detail according to the accompanying drawings showing a preferred embodiment of the present invention.

The present invention can be installed independently in the danger zone of the construction site, by analyzing the image of the worker approaching the danger zone to prevent disasters and accidents at the construction site using artificial intelligence image analysis that can guide the warning of the appropriate contents To provide, this invention includes a housing 10, a camera 20, a control unit 30 and a speaker 40, as shown in Figures 1 and 2.

The housing 10 is configured to be fixedly installed on the wall 1 or the like while the camera 20, the controller 30, and the speaker 40, which will be described later, are installed.

As shown in FIG. 1, the housing 10 is formed of a rectangular box-shaped enclosure and forms an accommodation space 11 of a predetermined size in which the control unit 30 and the speaker 40 are accommodated.

In addition, a plurality of fixing brackets 12 are fixed to the rear side of the housing 10 through fastening members such as anchor bolts (without reference numerals) to the wall surface 1 as shown in FIGS. 1 and 2, The front side is provided with a power switch 13 for turning the power on and off, and a USB terminal 14 and a LAN port 17 for moving data stored in the storage device 37 of the controller 30 to be described later.

In addition, the upper portion of the housing 10 protrudes a predetermined width toward the front side to block the power switch 13, the USB terminal 14, the LAN port 17 and the speaker 40 to prevent rainwater from entering. The member 15 is provided.

In addition, an installation table 16 having a predetermined length in the vertical direction is installed on the upper surface of the housing 10, and a camera 20, a speaker 40, and a warning light 50 are installed on the installation table 16.

At this time, the installation stand 16 is configured to be rotatable through a driving motor M or the like so as to be able to adjust the direction taken through the camera 20, and the camera 20 and speaker 40 are mounted on the upper part of the installation stand 16. And when the warning light 50 is installed, the direction of the photographing through the camera 20 can be appropriately adjusted while the installation table 16 is rotated at a predetermined angle by the operation of the driving motor M.

In addition, a power cable (not shown) for supplying commercial power to the camera 20, the controller 30, the speaker 40, and the warning light 50 is provided on one side of the housing 10, in addition to the housing 10 ), a built-in battery (not shown) is provided so that it can be operated continuously even if commercial power is not supplied for reasons such as a power failure.

At this time, the built-in battery is composed of a battery that can be used continuously for at least 2 hours or more, and is preferably configured to charge the built-in battery when commercial power is applied.

The camera 20 is installed on the outside of the housing 10 and is configured to photograph a dangerous area of a construction site, and the camera 20 is rotated in the vertical, horizontal, and horizontal directions to set a predetermined angle as shown in FIGS. 2 and 3. (A) The image in the range can be captured in real time, and in addition, it is composed of an infrared camera so that it can photograph workers approaching the danger zone even at night.

On the other hand, the camera 20, the speaker 40, and the warning light 50 are installed in a state exposed to the outside of the housing 10, and further include a waterproof function to prevent rainwater or the like from entering the inside.

The control unit 30 determines whether a worker, whether to wear protective equipment, or whether to access a dangerous area based on the image taken by the camera 20, and outputs a warning broadcast of the appropriate content through the speaker 40 if necessary. It is a configuration to be possible.

The control unit 30 recognizes the dangerous behavior from the worker's behavior pattern by repeatedly learning an image captured through the camera 20, and determines whether or not the protective equipment works through object recognition, for this purpose, TensorFlow ) The images input through the library are repeatedly learned, and multi-layer Perceptron-capable Keras is applied, and it is possible to learn multi-layer kernel measurement to recognize images, so whether workers and protective equipment are worn or not. A convolutional neural network (CNN) that is easy to classify is used. Here, Perceptron is a machine modeling the brain's learning function, and Keras is an open source neural network library.

Hereinafter, an example in which learning is performed through deep learning in the control unit 30 according to the present invention will be described.

First, for deep learning, using Keras' image data generator, a dataset necessary for initial learning is generated based on a general person's whole body image, and then the whole body image is divided into categories and repeatedly learned in Keras. Is ordered.

At this time, the whole body image is prepared by the number of wearable protective equipment such as a hard hat, a safety vest, and safety shoes, and the number of workers wearing the protective equipment.

After the data set is generated, features (images, animals, workers, etc.) for the video image are extracted based on the data set, and detailed classification is performed through a neural network based on the extracted features.

At this time, a convolutional layer is used to extract features from the video image. The convolutional layer is composed of a filter for extracting features and an activation function that converts the value of the filter into a nonlinear value.

In addition, since the filter of the convolution layer is implemented as a matrix, the input video image is also converted to a matrix, and then the result is obtained by multiplying the matrix value of the filter and the matrix value of the video image to derive whether or not it is a person.

At this time, the result value of the position other than the human face converges to 0, and conversely, the result value of the position corresponding to the human face comes out as a result value greater than 0. Through this, a feature is extracted by applying a filter to the previously prepared whole body image Then, this feature extracted image is made into an activation map, and then an activation function is applied to the created activation map, and the quantitative value produced by applying the activation function is non-linear indicating the presence or absence of the feature. The above process of converting to a value is repeated.

Here, as an activation function, a rectified linear unit (ReLu) is used, and includes 4 convolutional layers, 2 fully connected layers, and 1 dropout layer. do.

For more sophisticated performance improvement, only the front layer of the composite neural network (CNN) is imported and passed through to use the extracted features. ) Is excluded, and then the image data is put into this partial model and turned to contain the bottleneck features output from the last layer in an array.

At this time, all the bottleneck features consist of numerical values, and finally, a completely connected model is created to train this bottleneck data. In addition, data optimization is performed by performing a fine weight update on some upper layers of the previously learned network.

After learning is implemented by implementing a learning engine through the deep learning method as described above, a person (worker) can be detected from an image input through the camera 20, and whether or not protective equipment is worn from the detected person (characteristics) ) May be additionally detected, and it is possible to determine whether or not a person is approaching the danger zone by determining the route of movement of the detected person, so that, if necessary, a warning broadcast or warning light can be turned on and off.

In addition, the control unit 30 is provided with a storage device 31 of a predetermined capacity to allow the captured image to be stored for a predetermined time, and the data stored in the storage device 31 includes a USB terminal 14 and a LAN port ( 17) allows the administrator to copy and move them separately.

At this time, the login setting that requires a preset ID and password when accessing the control unit 30 through the USB terminal 14 and the LAN port 17 so that only the designated administrator can move or delete the data stored in the storage device 31 It can be included, thereby preventing an unspecified number from inadvertently acquiring data stored in the storage device 31.

Meanwhile, the warning light 50 may be configured to operate in accordance with the warning broadcast output through the speaker 40 under the control of the control unit 30. The warning light 50 is synchronized with the speaker 40 in the direction of the operator. It is configured to make it easier for workers to recognize danger by irradiating light sources or irradiating vulnerable areas.

Hereinafter, an example in which a disaster and accident prevention device at a construction site using an artificial intelligence image analysis according to the present invention recognizes a worker and performs a warning broadcast will be described.

4 and 5, when a manager such as a safety manager turns on the power of the disaster and accident prevention apparatus 2 installed in the danger zone 1A, the camera 20 installed outside the housing 10 Through this, an image in the vicinity of the danger zone 1A in a predetermined angle range is photographed in real time, and at the same time, the controller 30 determines the presence or absence of an object having motion from the image photographed through the camera 20.

Here, the camera 20 is pre-adjusted and installed at an appropriate angle so that a predetermined range of images including the dangerous zone 1A can be captured.

On the other hand, when there is an object having a motion in the captured image, the controller 30 analyzes the motion pattern of the object in order to determine whether the object is an operator. If and the range of the object with the movement is small, it can be judged by the movement of the object or the animal or the movement due to the diffuse reflection of light and the wind, not the operator.

When an object having movement is determined to be an operator through analysis of the motion pattern of the object, a unique ID of the object is given, and then whether or not protection equipment is worn is detected.

When a unique ID is assigned to an object determined as an operator in this way, even if the object moves in an irregular direction, it is easy to track, and in addition, many moving objects (workers) in the image because it is not necessary to additionally judge the object once determined. Even if exists, it is possible to quickly grasp the movement of an object and quickly process it.

Here, whether the object (worker) wears the protective equipment is judged by whether or not the feature in the input image is detected based on the data learned through deep learning.

In this process, if it is determined that the worker does not wear the protective equipment, a warning broadcast recommending the wearing of the protective equipment is controlled to be output through the speaker 40, and when it is determined that the protective equipment is worn, the object is moved. You will be tracking directions.

In addition, when the object being tracked (worker) approaches the danger zone 1A or the direction of movement is toward the danger zone 1A, a warning broadcast through the speaker 40 requires warning of the presence and caution of the presence of the danger zone 1A. It is controlled to be output, and in this state, it is further determined whether the object (worker) moves to the safety zone.

At this time, if the worker continues to move toward the danger zone (1A), the distance from the danger zone (1A) can be guided in detail in meters (M), and in addition, through this, the operator of the danger zone (1A) The position can be recognized more easily and quickly, and as a result, the operator can move to the safe area more reliably and quickly.

Here, when it is determined that the corresponding object (worker) moves to a safe zone (a position spaced at a distance greater than a predetermined distance from the danger zone 1A or outside the range of the image captured through the camera 20), the speaker 40 At the same time as the warning broadcast is ended, tracking of the object is ended.

Even if the object is tracked as described above, when re-entering within the range of the image captured through the camera 20 (or within a predetermined distance from the dangerous area 1A), a new ID is assigned to the object and then the process is repeated. It is done.

On the other hand, the object to which the unique ID is assigned is monitored by recording whether or not the protective equipment is worn. At this time, the value of 1 is stored when the protective equipment is worn, and the value of 0 is stored when the protective equipment is not worn. , When the input value is 0, a warning broadcast separately stored according to the corresponding data may be output. At this time, the protective equipment may be configured to add and delete monitoring targets for wearing or not wearing protective equipment as necessary, such as a hard hat, a safety vest, and safety shoes.

As an example of this, in a state in which the data entry table is configured to be stored in the order of the helmet and the safety vest, if the object does not wear the safety vest instead of wearing the helmet, the values of 1 and 0 are sequentially stored in the object, Through this, it can be judged that the corresponding protective equipment is not worn through the data of the safety vest with the input value 0. As a result, in the object tracking process, the warning broadcast for the protective equipment (safety vest) that the object is not wearing is a speaker. It can be guided through 40.

On the contrary, if a safety vest is worn instead of wearing a helmet, values of 0 and 1 are sequentially stored in the corresponding object, and thus, the corresponding protective equipment is not worn through data of the helmet having an input value of 0. As a result, in the process of tracking an object, a warning broadcast tailored to a protective equipment (hard hat) not worn by the object may be guided through the speaker 40.

In addition, when the worker wears the protective equipment through the warning broadcast in the surveillance area, the warning broadcast is ended with the corresponding input value being modified from 0 to 1.

Therefore, even if the number of protection equipments to be monitored is large, a warning broadcast tailored to the protection equipment can be output through the speaker 40, thereby enabling the worker to be more surely aware of the awareness of not wearing the protection equipment. It is possible to further induce equipment wearing.

As described above, the present invention is independently installed in the danger zone of the construction site, and can immediately output an appropriate warning broadcast to the worker approaching the danger zone without the control of the safety manager at the construction site, thereby improving the safety of the worker at the construction site. You will be able to secure more.

In the above, for convenience of description, the drawings showing the preferred embodiments and the components shown in the drawings have been described with reference numerals and names, but this is an embodiment according to the present invention. It should be construed that the scope of the right should not be interpreted, and the simple substitution from the description of the invention to a configuration that functions the same as the change to various predictable shapes is within the range that can be changed by a person skilled in the art for easy implementation. You will see it as self-evident.

1: Wall 1A: Dangerous Zone
2: Disaster and accident prevention device 10: Housing
11: Receptacle 12: Fixing bracket
13: Power switch 14: USB terminal
15: blocking member 16: mounting table
17: LAN port 20: Camera
30: control unit 30A: MCU
31: storage 40: speaker
50: warning light M: drive motor

Claims (5)

A housing 10 having an accommodation space 11 therein;
A camera 20 installed on the outside of the housing 10 to photograph an image in a predetermined angle (A) range;
A control unit 30 that is provided in the accommodation space 11 and analyzes images captured through the camera 20 in real time; And
A speaker 40 to which a warning broadcast is selectively output by the control of the control unit 30;
Including,
The control unit 30,
Based on the learning algorithm through deep learning, the image captured through the camera 20 is determined by whether or not the worker and the worker's protective equipment are worn through object recognition, and the danger behavior is recognized from the worker's behavior pattern.
TensorFlow library is used to iteratively learn images, Keras, which is capable of multi-layer perceptron, is applied, and Convolutional Neural Network (CNN) is used to recognize images.
Using the image data generator of the Keras to create a data set necessary for initial learning based on the image of the whole body and protective equipment of a general person,
When there is an object having a motion in an image photographed through the camera 20, the motion pattern of the object is analyzed to determine whether the object is an operator, and repetitive motion is performed while the object in the image is stationary. If it has or the range of movement is small, it is judged that the object is not an operator, but the movement of objects or animals or the movement caused by light reflection and wind, and if the object is determined to be an operator, a unique ID is given to the object. , Detects whether or not to wear protective equipment
When the object to which the unique ID is assigned continues to move toward the danger zone 1A, the distance separated from the danger zone 1A is guided in meters, and at the same time, a warning broadcast tailored to whether or not protection equipment is worn is provided to the speaker. It is controlled to be output through (40),
In the housing 10,
An installation table 16 having a predetermined length in the vertical direction; And
A warning light 50 for irradiating a light source toward the worker or the dangerous zone 1A in accordance with the warning broadcast output through the speaker 40 under the control of the control unit 30;
Is installed,
The mounting table 16,
It is configured to be rotatable through a driving motor (M) so as to adjust the direction taken through the camera 20,
On the mounting table 16,
The camera 20, the speaker 40 and the warning light 50 are installed, and the direction of the camera 20, the speaker 40 and the warning light 50 by the operation of the driving motor M Disaster and accident prevention device at the construction site using artificial intelligence image analysis, characterized in that the adjustment.
delete delete The method according to claim 1,
The control unit 30,
Disaster in the construction site using artificial intelligence image analysis characterized by extracting features for the input image using the generated dataset and classifying according to features through a neural network based on the extracted features And accident prevention devices.
The method according to claim 1,
The control unit 30,
After classifying the operator from the image received through the camera 20, additional judgment is made on whether to wear protective equipment, and a warning broadcast is selected according to the movement path of the worker and the wearing condition of the protective equipment, and the speaker 40 is used. Disaster and accident prevention device in construction site using artificial intelligence image analysis, characterized in that to be output.

Images