KR100789721B1 - A disaster prevention system and method - Google Patents

Abstract

The present invention compares the image of the construction site photographed by the camera with the previously stored image data to prevent safety accidents due to the wearing of protective equipment and safety accidents caused by various equipment and at the same time disaster accidents occurred on the construction site (fire, Regarding a disaster prevention system that can quickly determine the collapse, etc., the characteristic configuration is a camera 200 for photographing the monitoring target area 100 for monitoring the safety situation; Analyzes the current situation of the monitoring target area 100 using the image captured by the camera 200, reads the risk level according to the analysis, and if the risk level is high, transmits the content wirelessly and presents the current matter. Situational recognition device 300 to deliver from time to time; And a central antenna 400 provided with a server antenna 410 to wirelessly transmit and receive the situation recognition apparatus 300 to manage a disaster state of the monitored area 100.

Disaster, Prevention, Construction Site, Worker

Description

A disaster prevention system and method

1 is a reference diagram showing an example of a conventional civil construction site.

Figure 2 is a schematic diagram showing a disaster prevention system according to the present invention.

Figure 3 is a block diagram showing a disaster prevention system according to the present invention.

4 is a flowchart for explaining a disaster prevention method according to the present invention;

5 is a flowchart showing an operating state of a face recognition system.

6 is a video search example using a face recognition system.

7 is an exemplary diagram illustrating a face detection process using an SVM.

8 is a normalization process diagram through face region detection and feature point extraction;

9 is an illustration of a married couple area of a face.

10 is an exemplary view showing a result of a face recognition system.

11 is a schematic diagram illustrating a color-based image retrieval system.

12 is an exemplary diagram illustrating types of feature values of an image.

Fig. 13 is a graph showing the distribution form of the same sector image.

Fig. 14 is an illustration showing basic matching of NMFM.

15A to 15C are graphs of results according to the image measuring method.

16 is an evaluation table according to the image measuring method.

※ Explanation of symbols about main part of drawing ※

100: area to be monitored 200: camera

300: situation recognition device 310: control unit

320: image segmentation unit 330: image comparison unit

340: Image DB 350: field transceiver

360: Speaker 400: Central Server

410: antenna for the server

The present invention relates to a disaster prevention system, and more particularly, by comparing the image of the construction site photographed by the camera with the image stored in advance to prevent safety accidents due to wearing protective equipment and safety accidents by various equipment It is about a disaster prevention system that can quickly identify disaster accidents (fire, collapse, etc.) that occurred at the construction site.

Recently, interest in the prevention and management of accidents caused by various disasters is increasing rapidly. More than 100 people die each year from disasters, 50% of which are built. Civil engineering causes deaths due to industrial accidents, and 10% of them are due to fire and natural disasters. In addition to death, about 20,000 people are injured and seriously injured each year due to industrial and natural disasters. In addition to the loss of lives, our annual loss is more than 1 trillion won due to disasters.

On the other hand, the annual accident rate of phosphoric acid accidents increases by 5%, while the prevention and safety measures for such disasters are using the old methods used in the past (about 20 years ago).

In other words, due to the development of electronic and measurement technology, many fields use automated measurement management system using advanced devices and methods, but the field of natural disaster and business accident still maintains the system directly managed by the manager (human). Doing.

As such, the direct supervision by managers is not always complete due to the fact that it is not always on its own and is not supported by a lot of manpower. Therefore, a system for monitoring 24-hour disasters without manpower is urgently required. There is a situation.

Therefore, many fields have been researching such system development. Various methods have been provided, including the technique of using a sensor to notify the fire, and a device that attaches the sensor to the leg before the bridge collapses. However, such a disaster prevention system was designed and developed for a specific field and a specific environment, and thus there was a limit in preventing various environments and various types of disasters.

On the other hand, the best way to prevent disasters is for humans to understand the environment with their own eyes, to see the contents and to detect dangers. However, it is environmentally very difficult for humans to manage risks at all times.

Recently, as shown in FIG. 1, the current situation is photographed through a camera, and the captured image is remotely recognized by the administrator and the risk of the matter is detected. A system for reporting to managers has been developed and used.

FIG. 1 shows an example of a civil construction site, in which a crane 1 is constructing a hill 2 in which two workers 3 are not in a safety zone and are wearing a safety helmet and hat. Form.

If a dangerous situation occurs in the crane (1) in the upper part and the stone or dangerous goods hit the worker (3) is subjected to a very large industrial accident accident without wearing safety equipment.

On the other hand, the manager who checks the current situation information through the camera (4) as an image to inform the dangerous situation to the operator (3) or the crane (1) driver in the field through the speaker (5), which can prevent the disaster will be.

However, even at this time, there was a problem in that it depends on the administrator who can check and take action on the site through video to prevent disasters.

There are also many environmental hazards during construction and civil works. There are several safety mechanisms and mechanisms to work safely from these hazards, but the loss of life is caused by the failure to utilize them.

And even though workers must wear helmets and safety shoes in the work area, they can wear a helmet and safety shoes and cause serious injuries in the event of an accident, and there will be unauthorized vehicles and workers in explosive areas such as tunnel construction. Personnel accidents can occur and landslides and other disasters can occur in construction areas.

On the other hand, the damage caused by forest fires is increasing year by year. If a fire can be known in advance, it can be prevented in advance, but the currently used forest fire monitoring system checks the image taken by the camera as described above, and has a limitation in making a quick and accurate judgment in case of a fire. There was this.

The present invention has been invented to solve the above problems, the object of which is to recognize the state of the video taken by the camera to recognize the set risk factors and inform the central server of the contents.

That is, the present invention divides the image taken by the camera into a predetermined size and compares the divided image with the image data stored in the image DB to prevent accidents at construction sites that can be notified to workers and managers in case of safety accidents or disasters. It is to provide a prevention system.

Referring to the characteristic configuration of the present invention for achieving the above object is as follows.

Characteristic configuration of the disaster prevention system of the present invention is a camera for photographing the monitoring target area to monitor the safety situation; Analyzing the current situation of the monitored area by using the image captured by the camera, and reading the risk level according to the analysis contents, and when the risk level is high, wirelessly transmits the contents and transmits the current matter from time to time. ; And a central server provided with a server antenna to wirelessly transmit and receive the situation recognition device to manage a disaster state of the monitored area.

In addition, the characteristic configuration of the disaster prevention method of the present invention comprises the steps of photographing a surveillance target area to monitor the safety situation with a camera; Dividing the image photographed in the step into predetermined sizes; Determining the disaster situation of the monitoring target area by comparing each image divided in the above step with the image data stored in the image DB, and in the step, notifying a worker or manager of the field by voice through a speaker when the disaster situation occurs. It was done.

Referring to the present invention having such characteristics in detail as follows.

The present invention is a disaster prevention system that includes an image pattern recognition technology for detecting a danger when a disaster occurs using the image recognition technology.

Figure 2 is a schematic diagram showing a disaster prevention system according to the present invention, Figure 3 is a block diagram showing a disaster prevention system according to the present invention.

As referred to herein, a camera 200 capable of capturing a surveillance target region 100 to monitor a safety situation is provided, and the camera 200 uses the photographed image to present the current of the surveillance target region 100. The situation is analyzed and the risk level is read according to the analysis, and when the risk level is high, the contents are notified to the center through mobile communication and the current situation is frequently connected to the situation recognition device 300.

The situation recognition apparatus 300 is provided with an image splitter 320 for dividing an image taken by the camera 200 into a predetermined size, and an image for comparing with the image split by the image splitter 320. An image DB 340 is provided to store data, and the image DB 340 may be present in an area to be monitored 100. Image data corresponding to objects and phenomena (fire, collapse, etc.) are stored.

In addition, an image comparator 330 for comparing the image data transmitted from the image DB 340 and an image divided by the image splitter 320 is provided, and a controller 310 for controlling the operation of each part is provided. In addition, a field transmitting and receiving unit 350 for transmitting a signal generated by the control unit 310 and receiving a signal transmitted from the outside is provided.

On the other hand, the central server 400 is to check and control the operation status of the plurality of situation recognition apparatus 300, the central server 400 for the server antenna for transmitting and receiving data with the situation recognition apparatus 300 (410) ) Is connected.

4 is a flowchart illustrating a disaster prevention method according to the present invention.

As described herein, a step of photographing the surveillance target area 100 to monitor the safety situation with the camera 200 is performed, and each of the images photographed in the step is divided into predetermined sizes.

The divided image as described above is compared with the image data stored in the image DB (340) to check the wearing state of the protective equipment of the worker working in the monitoring area 100 and at the same time determine the operation of the equipment and the occurrence of disaster Perform.

If there is an operator who does not wear the protective equipment in the step is to notify the worker wearing the protective device to the voice through the speaker 360 installed in the field by voice.

On the other hand, when an abnormal state or a disaster (fire, collapse, etc.) of the equipment in the above step is notified to the operator through the speaker 360 and at the same time to transmit the emergency situation through the transmission and reception unit 350 for the field.

The signal transmitted in this way is received by the antenna for the server 410 is transmitted to the central server 400, the central server 400 notifies the administrator of the emergency and at the same time the emergency measures (safety personnel, It will be possible to promptly handle safety accidents and disasters by notifying emergency situations to the fire department.

The present invention has the following technology.

1. A human state recognition technology that recognizes a person's face and recognizes a person's actions and the helmet or boots he currently wears.

2. When a fire occurs in a specific space such as a mountain or a workplace, a technology that recognizes the fire and informs the contents through a mobile.

3. A technology that recognizes the working radius of the excavating machine, etc., if there is a person or other equipment within the radius, notifies the danger, records the contents, and sends it to the server via mobile if necessary.

4. Technology that recognizes license plates and confirms and manages whether the vehicle is accessible.

5. A technology that defines patterns for certain hazards through other cameras and sends them to the center when those patterns occur.

On the other hand, due to the development of computer performance, network, data compression technology, etc., the amount of multimedia data is increasing tremendously, and there is a need for technology for managing / using multimedia data such as content-based video retrieval and video indexing. Automated cut detection is a starting point to meet these various needs.

The images inputted by the camera are stored as a single video material, and a technology for processing and managing the video material is required. The minimum unit constituting the video is a frame representing one image. A sequence of frames continuously obtained by one camera is called a shot, and a set of shots that are related in content is called a scene. Cut refers to a scene change point corresponding to a boundary between the scene and the scene. In order to accurately analyze the current situation, it is necessary to have a technique for separating and processing scene-related cuts and scenes in the current video flow.

To date, various methods for cut detection have been proposed. These methods can be classified into two types, one for uncompressed video sequences and one for compressed video sequences. Methods applied to uncompressed video sequences include comparing pixel differences between two consecutive frames, comparing histograms, and comparing motion vectors. Methods that can be used for a compressed video sequence include a method of comparing compression coefficients and a method using I, P, and B frames.

Even if the cut is found automatically, it is difficult to use the detected cut as it is for applications such as video indexing. Multiple cuts can exist within the same shot, since the unit of video that humans see is meaningful is the shot, not the cut. Therefore, there is a need to store the detected cuts hierarchically and structurally according to each shot.

The cut detection method is as follows.

In the present invention, color histogram and edge information are used. The change in the histogram and the edge information is observed between two consecutive frames. If the change exceeds the given threshold, the cut is determined.

The HSV color model is used to calculate the color histogram. The HSV color model has the advantage that the method of expressing color is similar to the method of human recognition, compared to the RGB color model, and thus the distance between colors is more similar to that of human being.

H, S, and V components are used in the color histogram, and each histogram section is quantized properly. This is to reduce the amount of computation, but also because the quantization of the histogram is less sensitive to small changes. The quantization levels of the H, S, and V components were determined differently depending on how important humans feel. 18, 12, and 16 levels are used, respectively.

The sum of the amount of change in each color histogram between two consecutive frames M and M-1 was defined as DC (x, y), as shown in Equation (1).

[Equation 1]

(for each H, S, V)

: M번째 프레임의 i 번째 히스토그램 값

: I th histogram value of Mth frame

The Sobel operator is used to extract edge components within a frame. After the Sobel operator is applied to a gray image, the number of pixels having gray values above a threshold is obtained, and the difference is used as a feature value for cut detection in two consecutive frames.

[Equation 2]

S (x, y): Sobel boundary image

The difference between two consecutive frames is obtained by Dist. Color and Dist. It is determined by the combination of edges. If this value is over the threshold, it is determined that a scene change has occurred.

On the other hand, the use of global information is described as follows.

In order to classify each detected cut for each scene, a method of using global information is first examined. The method used to calculate the difference between two cuts uses the same method as the one used for cut detection. That is, the difference between the two cuts is calculated using the color histogram and the edge information of the two consecutive cuts already detected, and if the value is greater than or equal to the threshold, it is determined that the two cuts are included in different scenes.

In addition, when using local information, in order to use local information for comparison between cuts, two cuts to be compared are first divided into 3x3 sections, and the same feature values as those used for cut detection are extracted for each block. Bipartite matching is used to compare the feature values. Bipartite matching finds a combination that best matches each node in a bipartite graph.

After extracting the histogram and edge components of each block of the previous and current cuts to be compared, the difference is calculated for each block, and a Bipartite graph having the weight as the difference is constructed. By using Bipartite matching in this graph, we can obtain the combination that minimizes the difference between the two cuts. The difference between the two cuts through this combination is used for scene detection.

In addition, the worker authentication and situation analysis technology using face recognition technology requires a technology of detecting a face on the input received by a camera and recognizing the face by introducing a face recognition technology into a safety management system to know who the worker is.

As shown in the face recognition flowchart of FIG. 5, a face of a person may be detected from an image, a face may be recognized and indexed, and a face may be quickly recognized from a later input image. Figure 6 shows an example of such a function, it is possible to check the wearing state of the safety device of the worker with the face using this technique.

In addition, the present invention uses a SVM (support vector machine) as shown in Fig. 7 for face detection. In order to detect a face regardless of rotation and size using SVM, the size and rotation angle of a mask scanning an image must be changed. Scanning the entire area of the image and detecting a face like this greatly increases the computation time. In order to solve this problem, a method of detecting a candidate face in the entire image and then detecting a correct face area in the candidate area is used. Normally, skin color information is used to detect a face candidate area. When detecting a candidate area by flesh color, there is a problem in that the candidate area is not properly detected in an image in which illumination is severely changed or a black and white image. In this system development, we will develop a complete face recognition system by conducting in-depth research.

And, face recognition is a very difficult problem, which suggests that one person's face change is greater than the other's face difference. Data that is processed from the computer's point of view by a person yawning, closing eyes, hair style, beard, wearing glasses, aging, camera shooting position, noise, lighting, etc. As a result, facial changes in one person can be much greater than facial differences in another person. In order to solve this problem, it is necessary to perform calculations such as correction according to the photographing position of the face image or noise removal through a preprocessing process. In order to solve this problem, this study performs accurate localization of face image by predetermined rule. The main method for registration of the current face is performed by detecting the positions of both eyes as shown in Fig. 8, and then correlating the detected eye positions with predetermined coordinates. This requires accurate detection of both eyes.

The eye detection method can be divided into low level image processing method and statistical method. When using low level image processing, it is possible to use the shape of eyes or pupils using edges or information on intensity itself. This method has a disadvantage in that it is applied to a relatively high resolution image. In this study, we can approach the statistical method using SVM to detect the eyes in low resolution images.

In addition, face recognition has long been studied in computer vision studies, and has been used for face database applications such as security, police investigation, and personnel management. Face recognition first detects whether or not a face exists from an input image and normalizes it using feature points such as eyes. Feature vector information is extracted from the normalized face image, and the feature vector is compared with the feature vectors of the face images created through the same process to recognize who the face is. The facial feature vectors may be extracted from the entire face image, and the facial images may be divided into partial regions such as eyes, nose, and mouth, and the feature vectors may be extracted in units of partial regions. An example of segmentation of the face region in the case of extraction into partial regions is shown in FIG.

As referred to herein, the entire face area is divided into three partial areas centered on the eyes, nose, and mouth. In the present invention, both methods may be used or either may be used.

To develop a face recognition system, defining a feature space that can express facial features is an important key. Currently, many studies have adopted face recognition technology using Principle Component Analysis (PCA). As shown in FIG. 10, PCA develops and uses the actual system by analyzing the most common characteristics that appear in the database for the face images in the database and defining them as feature values representing each face. . However, the PCA has similar characteristics in each image due to its characteristics, and thus, the recognition rate is remarkably reduced when there are many faces to be recognized. The present invention has recently been applied to the field of face recognition-Linear Discriminant Analysis (LDA), Independent Component Analysis (ICA), Non-parametric Methods, Neural Network, Baysian Classifier, HMM (Hidden Markov Model) By using-, strong face recognition is possible.

In addition, the present invention can be largely divided into three types of hardware manufacturing technology that can recognize multimedia data input transmission technology and image pattern recognition technology.

Also, in image pattern recognition technology, each content such as face recognition, situation recognition, and shape recognition belongs to advanced technology.

And the present invention has four features as follows.

First, an image pattern recognition method capable of recognizing a situation uses a near maximum flow matching method using an approximate maximum flow. Using this method, the image pattern can be recognized and read out in real time.

Second, the face recognition part uses the current PCA method and the matching method uses the WBM method to correctly recognize the face.

Third, the mobile network technology works in conjunction with the CDMA technology to optimize and compress video information for high speed transmission.

Fourth, the hardware configuration is the state-of-the-art technology to construct the lowest cost hardware using the lowest cost. Making a low-cost system capable of maintaining optimal functionality is part of the excellence and ingenuity of the technology.

Hereinafter, image pattern recognition will be described.

The image pattern recognition technology of the present invention evaluates the superiority of the technique we have in comparison with the types of image pattern recognition and existing researches.

Types of Image Pattern Recognition System

Existing image recognition system can be divided into index-based image recognition system and content-based image recognition system. Index-based research has been conducted in the early stages of research on image retrieval, and has been used for most commercialized systems until now.

Index-based image recognition system is a method using a text string that indexes the information of the image with the image. 11 shows an example of an index-based image recognition system.

The index-based image recognition system has a merit of finding exactly the desired search because it is a method of accurately inputting a feature and information of an image as a string. However, in order to fit various searches, information about images must be variously indexed.

For example, to find an image that matches three pieces of information: "bear", "running", and "baby bear", you need to index the image in several ways. This takes a lot of time and effort to build the database. This technology is not suitable for use in systems that recognize patterns of objects in real time.

The content-based image recognition system is a system that automatically extracts feature values using database contents and searches them by database. The MPEG-7 standard "Multimedia Content Description Interface" of ISO / IEC 15938 defines various descriptors that represent feature values of video data. This descriptor is intended to standardize various information included in moving image as well as still image, and the purpose of standardization is to share effective data information in retrieval of contents.

The visually expressed part of the MPEG-7 standard is called a visual part and is referred to as a visual part in the present invention. The visual part of MPEG-7 defines descriptors that can analyze various types of image features. Not all descriptors in the visual portion of MPEG-7 are actually descriptors used for searching, and some descriptors list only attribute information serially from the display medium. Some of the descriptors are also collections or parts of descriptors. Basic descriptors include Color Layout, Color Structure, Dominant Color, Hierarchical Color, Edge Histogram, Homogeneous Texture, and Texture Browsing. Browsing, Region Shape, Contour Shape, Camera Motion, and Motion Distribution.

Systems using descriptors are based on low level descriptors that support the system.

GOF / GOP (Group-of-Frame / Group-of-Picture) uses hierarchical color descriptors, 3D shape information construction uses 3D mesh-information, and Motion Trajectory is an object. Segmentation information is used, and Face Recognition uses face extraction information. Descriptors and localizations for spatial combining are constructed using information such as specific 2D coordinates, grid layout area allocation, and time series.

MPEG-7 visual descriptors are characterized by color layout descriptors (CLD), color structure descriptors (CSD), dominant color descriptors (DCD), and hierarchical color descriptors (SCD). There is.

Descriptors for texture retrieval include Edge Histogram (EHD), Uniform Texture Descriptor (HTD), Texture Browsing Descriptor (TBD), and Shape / Contour Based Shape Descriptor (RSD / CSS). Descriptor information is defined by the standard of MPEG-7. In the information extraction process, each descriptor extracts the defined parameter in the whole image. For example, the color of the color structure uses 32 quantized bins, the effective color extracts feature values from the YCrCb color space, the uniform texture descriptor is quantized into 32 elements, and the feature values of the hierarchical color descriptors. Uses the 3-bit value as tag information and 64 quantization feature values.

The application and the advantages and disadvantages of the image pattern recognition system will be described as follows.

Existing image recognition systems are used in various fields. To date, index-based retrieval systems have been dominant, and content-based image recognition systems are gradually increasing. Index-based search system has the disadvantage that the search section is limited, but there is an advantage that can be searched correctly in the limited search section.

For example, if the information of the image is classified by the animal name, the animal name can be searched correctly. However, the search is difficult when searching by the movement or property of the animal.

In addition, index-based technology cannot be used to read safety status in real time. Because content-based image recognition systems are based on color and shape patterns, they are often used for similar retrieval rather than classification.

The disadvantage is that there is no index classification, so the exact search is not performed according to the classification. For example, when searching for information in the form of a tiger, all images having the color and pattern of the tiger can be searched, but the image may not be limited to the tiger.

Index-based image recognition requires a lot of time and effort to index the information, and it is difficult to construct a variety of information of the image. In content-based image recognition, a system using a single information is mainly used to recognize an image. When retrieving an image using only single information, if single information is changed or there is no stable provision of single information, accurate search cannot be performed. In addition, since the search using the descriptors proposed in MPEG-7 is due to standardization, the performance of the standardized information is excellent, but it is difficult to search in a special system that cannot use the standardized information. Therefore, there was a problem that a proper method should be suggested in the special type of search.

The necessity of a content-based image recognition system is as follows.

Problems of the index-based image retrieval system and the content-based image retrieval system have been described above.

That is, the biggest problem of the index-based image recognition system is that indexing is required according to the image data, and it takes a lot of time and effort to make an image of an exponentially increasing image database. Indeed, many systems are currently shifting from index-based to content-based to overcome this limitation.

Content-based image recognition can be divided into a system using a single information and a system using multiple information. The single information refers to a system for searching by using information of colors, shapes, and patterns, and the multiple information refers to a system combining colors and shapes, colors, and patterns. While single information minimizes the amount of information, the characteristics of the image are dependent on one information, so large changes in the information cannot accurately search. In order to solve this problem, a content-based image retrieval system using multiple information has been developed in many fields.

As shown in FIG. 12, a recognition system using a single feature may be regarded as color, shape, and pattern as feature values representing a feature of an image. In general, the most commonly used color when searching for images such as characters. In addition to the character image, the form is used for a trademark search or an object search with a small change in form.

When using shape information, two representative information is used. Region shape information and contour shape information.

Area-based shape descriptors typically extract feature values using a basis function such as moment. Typical domain-based shape descriptors are Zernike Moments or Descriptors using ART. Retrieval system using shape information is mainly used in static image retrieval-system without change of shape such as trademark retrieval system or character retrieval and logo retrieval.

Image retrieval using color information has been actively conducted in recent years. When using color information, color histograms are generally used. Swain's method of indexing using color histogram is proposed. This method creates a color histogram of an object and quantizes the histogram to extract feature values. The Swain method has a light-sensitive disadvantage. Brain proposed a method of quantizing a color histogram that is not related to lighting. Many search systems using color histogram quantization values have been proposed using this method, and Gevers's proposed method is the most recently used method.

 When using a color histogram, it is common to use a histogram of three RGB regions. Swain transforms the color histogram into regions with three RGB feature values correlated with black-white, red-green, and blue-yellow. The information value is compressed to 2,048 bins by allocating 8 bits per channel of the region to be switched. Using this method, you can search the information of the image very quickly. The images in the database have a color histogram of the object image that excludes the background, and store the 2,048 bins of the color histogram. In other words, when the feature value of the color histogram is extracted from the query image to be searched and compared with the image in the database, the number of bins is small so that the comparison can be performed very quickly.

In addition, the retrieval system using multiple features recognizes images using color and shape information at the same time. Gevers proposed a system for recognizing images using unchanged color and color information. The method proposed by Gevers uses the shape information of the object as geometry information and the method of storing and retrieving the color information of the corresponding area. Using colors and shapes increases the amount of information than retrieving a single piece of data.

Therefore, it is essential to optimize the amount of information when using multiple pieces of information. The method proposed by Gevers prioritizes the accuracy of information retrieval over optimization of information volume. When the image is searched using the color information and the shape information together, the accuracy may be higher or lower than the method of searching the information of the color or shape in a single way. The reason for the lower accuracy is that the same image is not processed in the same image if the error value of one of the colors or the shape is large. This is because when two pieces of information are viewed at the same time, the error may be greater than that derived from one piece of information. The method that Gevers suggests as an invariant value of information even while viewing color and shape information at the same time.

As described above, various methods of image recognition technology have been developed. The most important aspect of image recognition is whether or not to use multiple information.

For example, in order to recognize whether a person wears a hard hat, the method of recognizing the shape of the hard hat and the color of the hard hat is superior to the method of recognizing only the color information or the shape information of the hard hat. The image recognition technology of the present invention is a method of measuring and recognizing similarity distribution using multiple information. By using this method, the situation can be recognized more accurately than the existing image recognition method.

The recognition method of the present invention is a flexible recognition method in which the recognition data is not fixed. Therefore, if the recognition element is changed, the corresponding part can be recognized quickly. For example, by providing recognition information about the hard hat, the hard hat is recognized, and if the information of the fire situation is set as an element, the dynamic recognition can be performed to recognize the fire.

를 구하고, 질의 영상과의 비교 값이

이하 일 때 같은 영상으로 처리한다. 대부분의 영상 비교가 통계적인 방법을 이용하고 있으며, 판별을 요구할 때 통계의 의해 결정지어진 특정 문턱치값

를 이용한다. In the present invention, when extracting and comparing multiple features in an image object, a threshold value is extracted using a statistical method, and similarity is measured using this value. When using the statistical method, as shown in Fig. 13, the distance difference distribution between the images is obtained through a comparison experiment of a large number of image data, and then the specific distance difference when the image is mostly composed of the same image.

And compare with the query image

The same image is processed when Most image comparisons use statistical methods, and specific thresholds determined by statistics when requiring discrimination.

Use

값을 구해야 한다. 값이 없을 경우 WBM은 단순히 유사도가 큰 순서로 영상들을 정렬 할 수밖에 없다. 영상 검색의 목적은 영상의 유사도 정렬이 아니고 같은 영상, 또는 유사도가 특정값 미만인 영상을 검색하는 것이다. 따라서 WBM을 이용하여 영상을 비교하는 경우에도 실험을 통해

값을 구해야 한다. When comparing images using WBM, experiments are conducted to distinguish the same image.

You need to get the value. If there is no value, WBM simply sorts the images in order of similarity. The purpose of the image retrieval is not to sort the similarities of the images, but to search for the same image or an image having a similarity less than a specific value. Therefore, even when comparing images using WBM,

You need to get the value.

Until now, the method has been used to determine whether the value is the same image after the comparison result of the image, but the proposed NMFM is characterized by using the value as a prediction index. In general, the similarity distribution of an image is used to extract a specific threshold using a statistical method. Equation 3 is a general similarity distribution equation.

[Equation 3]

는 질의 영상인 와 데이터베이스 영상인

와의 거리차를 의미한다. 총 n개의 영상을 모두 비교하였을 때

값의 거리차가 있는 수를

로 표시하고, 전체적인 분포는 값의 총합을 의미하며 이 값이

이다.

Are query images and database images

The distance between and When comparing all n images

The number of distance differences

, The overall distribution is the sum of the values,

to be.

Fig. 14 is a distribution curve obtained using equation (3). Determining the standard deviation after the actual experiment depends on the system.

로 정한다. NMFM에서는

를 Bipartite Graph에서 더 이상 자라지 않는 노드 설정에 이용한다.For example, if the system requires 95% accuracy, the end point of the area where 95% of the same image is

Decide on In NMFM

Use to set up nodes that no longer grow in the Bipartite Graph.

는 최소 흐름을 가지는 쌍이며 특정 섹터에 연결된 가중치 값들이다.When setting the optimal pair in the minimum flow matrix, N edges are set, the minimum cost of each edge is set in the matrix, and the pair having the minimum cost is obtained. The method of finding the minimum flow is shown in Equation 4.

Are the pairs with the minimum flow and are weight values connected to a particular sector.

[Equation 4]

Once the minimum flow pair is determined, the pair is removed and the next pair is obtained by repeating Equation 4. Find the pair, then get the total flow and see it as the approximate maximum flow.

값 보다 클 경우, NMFM에서 WBM을 이용하여 최적화 흐름을 유도한다. 실험을 하는 동안 같은 영상이면서

값보다 큰 경우가 발생하지 않는다. 같은 영상일 경우 2500개의 영상에서 모두 한 개 이상은 존재 하는 것이다.In the present invention, if the weight of all pairs is

If greater than this value, the NMFM uses WBM to derive the optimization flow. During the experiment, the same video

It does not occur if it is larger than the value. In the same image, at least one of the 2500 images exists.

The disaster prevention method by the NMFM described above is as follows.

의 차이 값을 구한다. First, all the connections that can be connected to the image data are set in the image divided into a predetermined size, and at this time, a pair for obtaining the minimum cost is extracted. Where the weight of the pair

Find the difference between.

의 차이가 소정의 범위(영상 데이터의 유사범위)보다 클 경우 NMFM에서 WBM 알고리즘을 이용하여 최적화 되는 흐름을 구한다.Obtained in the above step

If the difference is greater than a predetermined range (similar range of the image data), the NMFM calculates the optimized flow using the WBM algorithm.

On the other hand, if the difference of σ obtained in the above step is smaller than a predetermined range (similar range of the image data), a connection pair is obtained by comparing with another image data (extracting image data similar to the captured image).

As described above, when continuously connecting pairs are set, the coefficients of the edges that can be ignored because they do not exceed the error range are set and the pairs exceeding the coefficients are sequentially ignored.

Then, the flow value is divided by the value of the connected edge to obtain an average flow value.

값과 비교하였을 때 해당 구간 안에 있을 경우는 도15a 내지 도15c에 나타낸 방법을 수행하고,

의 범위를 벗어났을 경우 또는 완전한 정합 시

구간에 있는 영상이 하나도 없을 경우 대략적 흐름에서 매우 큰 흐름을 유지하기 때문에 같지 않은 영상으로 처리할 수 있다.15A to 15C are measurement graphs for explaining the basic method of NMFM according to the present invention;

When compared with the value, if it is in the corresponding section, the method shown in Figure 15a to 15c,

Out of bounds, or on full match

If there is no image in the section, it can be treated as an unequal image because it maintains a very large flow in the approximate flow.

값의 영역 내에 있는 영상이 없을 경우, 더 이상 계산을 하지 않고 전혀 다른 영상으로 처리함으로써 계산을 하지 않고 비교하면서 빠르게 연산할 수 있다. 두 번째로 모든 영상에서 거리차를 계산하여 정렬 하고자 할 경우

안에 있는 영상들은 NMFM을 이용하여 연산함으로써 시간을 절약하고 나머지 부분은 WBM을 이용하여 연산하여 유사한 영상의 정렬을 할 수 있다. As such, NMFM can save time in two ways. at first

If there is no image in the range of values, it can be computed quickly by comparing the calculations without further calculations by processing them as completely different images. Second, if you want to calculate distance difference in all images

The images inside can be saved using NMFM to save time, and the rest can be calculated using WBM to align similar images.

In the basic registration of the NMFM, when the sector image is placed at the edge, a very large distance difference appears in a specific bin in the image having a large change. The value of this distance difference appears as a large error value in the image similarity. The NMFM ignores the sector image where the distance difference is very large and maintains the maximum flow with only the remaining sectors. Equation 5 is to calculate the maximum flow.

[Equation 5]

는 무시하고자 하는 섹터의 개수를 의미한다.From above

Denotes the number of sectors to be ignored.

The characteristics of such NMFM are as follows.

The big difference between NMFM and WBM is whether to extract matched pairs. In addition to the maximum flow chart, WBM also finds matching pairs when the maximum flow chart is achieved. However, NMFM extracts the maximum flow rate but does not find a matched pair. The purpose of image matching is not to find matching pairs, but to find the maximum flow chart. The computational complexity is much smaller than that of WBM because it eliminates unnecessary operations for finding matching pairs and predicts flow diagrams like WBM. Using the WBM algorithm for N vertices, the fastest way is to do n-1 traversals. NMFM is always equal to or less than the operation of WBM because when the flow is less than or equal to the value of n-1 traversals of NMFM, no further traversal is necessary. Therefore, the operation of NMFM is faster than WBM. Fig. 16 shows the associative pattern recognition result using the present invention.

In content-based retrieval systems, evaluation scales are important indicators of the accuracy of the system. In the present invention, in the character search, evaluation scales were used as BEP (Bull's Eye Performance), NMRR (Normalized Modified Retrieval Rank), and NDS (Normalized Distance Sum). BEP is a measure of the shape and motion descriptors of MPEG-7, and NMRR is a measure of the color and texture descriptors of MPEG-7. NDS has the advantage that the closer to 0, the better the performance, and all the evaluation conditions are considered and detailed measurement is possible.

The smaller the numbers, the better the NMRR and NDS. The closer to 1, the better the BEP. WBM and NMFM are far superior to the commonly used GHM and PicToSeek methods.

As described above, the present invention compares real-time information captured by a camera with image data stored in an image DB in real time to prevent safety accidents due to carelessness of field workers, as well as to quickly obtain information when a disaster occurs, thereby minimizing damage caused by a disaster. There is a unique effect that can be done.

Claims (5)

A camera 200 installed in each of the surveillance region 100 to photograph the surveillance region 100 to monitor the safety situation; An image splitter 320 for dividing the image photographed by the camera 200 into a predetermined size, an image DB 340 for storing image data for comparison with the image split by the image splitter 320, and After measuring the similarity distribution so that the accuracy of information retrieval takes precedence over the optimization of the amount of information so that the disaster situation determination time of the image divided by the image dividing unit 320 is minimized, the information of the situation is set as an element and the image DB 340. The image comparator 330 for comparing the image stored in the image and the divided image, the controller 310 for controlling the operation of each part, and the speaker 360 for transmitting the audio signal generated by the controller 310 to the worker And a field transmitting / receiving unit 350 for transmitting a signal generated by the controller 310 and receiving a signal transmitted from the outside, to be monitored using the image photographed by the camera 200. Analyzes the current situation of the station 100 and reads the risk level according to the analysis, and when the risk level is high, the situation recognition device connected to each camera 200 to transmit the contents wirelessly and deliver the current matter from time to time ( 300); And a central server (400) provided with a server antenna (410) to wirelessly transmit and receive the situation recognition device (300) to manage a disaster state of the monitored area (100). delete delete Photographing a surveillance target area 100 to monitor a safety situation with a camera 200; Dividing the image photographed in the step into predetermined sizes; The similarity distribution is measured so that the accuracy of information retrieval takes precedence over the optimization of the amount of information so that the disaster determination time of each image divided in the above steps is minimized, the situation information is set as an element, and the images stored in the image DB 340 are compared. Determining a disaster situation of the monitoring target area 100; And notifying a worker or a manager in the field by voice through the speaker 360 when a disaster situation occurs in the step. delete

Images