JP2006031645A - Real-time estimation method for dynamic crowd density and crowd accident prevention system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide real-time estimation technology for dynamic crowd density based on image processing for caution and prevention of generation of a crowd accident in a store, various kinds of event sites or the like. <P>SOLUTION: In this real-time estimation method for dynamic crowd density and crowd accident prevention system, so as to grasp a congestion situation in real time to contribute to crowd control, the number of persons is not accurately counted by closely watching the individual person, but a property as a crowd is characterized into a moving lump or a lump having a different background and different brightness, rigidity is not pursued, and simplification is performed as thoroughly as possible to a degree not impairing practicality. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an image processing technique that warns and prevents the occurrence of an accident in a store or various event venues.

  At festivals, event venues, and new store opening venues where many people are crowded, an unspecified number of people and cars are concentrated and crowded. The duties of the hustle guard are to prevent traffic accidents and confusion caused by hustle and buses by appropriately guiding visitors to traffic and arranging entrances and exits. In recent years, an accident report (Non-Patent Document 1) of the Akashi Festival (fireworks display) accident in Hyogo Prefecture caused an accident that occurred, although it was possible to predict that the pedestrian bridge at the accident site would be a bottleneck. He argued that there was a problem in not having prepared a countermeasure in advance and that the accident could be prevented by restricting the inflow to the pedestrian bridge. The number of people staying on the pedestrian bridge is calculated from various data at the time, and the timing of inflow restriction is also described, but there was no system that could respond appropriately, and such a system was naturally installed. It wasn't.

Akashi Civic Summer Festival Accident Investigation Committee: Fireworks display accident investigation report at the 32nd Akashi Civic Summer Festival (2002) http: // www. city. akashi. hyog. jp / soumu / h safety / natsumatsuri hokkoku. html

  Knowing the situation where the number of people increases over time is important for judging the timing of regulation. This timing is difficult due to crowd rebound after the crowds get worse. By tracking changes in the number of people over a certain period of time, it seems that objective and effective timing can be obtained. Because the fireworks display accident investigation report (2002) in the 32nd Akashi citizen summer festival suggests that it is difficult to respond flexibly immediately after the occurrence of the accident due to congestion, It is important to implement measures to pick buds in advance in a timely manner.

  In the present invention, a technology for grasping the congestion situation in real time and contributing to crowd guarding has been realized. Unlike the existing technology, the technology of the present invention does not count the number of people accurately by paying attention to each person, but has the originality in characterizing the characteristics as a crowd. In addition, it is a technique that does not pursue strictness and simplifies as much as possible without impairing practicality, and determines necessary conditions.

  There are the following three technologies as reported / existing technologies related to the present invention.

  The first is “Real-time measurement of pedestrian distribution in town and creation of maps” (Non-Patent Document 2) described in “Koji Maeda, Kunihiro Chihara”, Master's thesis of Nara Institute of Science and Technology (2002). is there. The technical contents are as follows. Pedestrian distribution information is also useful city information that has its own purpose of use, such as being able to be used as a material for studying marketing and urban development plans. We have proposed a system that can extract pedestrian distribution information in the city and provide it to users. The pedestrian distribution measurement uses image processing capable of unspecified many sensing, and uses a camera equipped with a fish-eye lens capable of photographing a wide range. Pedestrian distribution measurements need to be performed at multiple locations, but the information extracted by using the network is collected at one location, enabling integration on the map. Whereas this existing technology uses a plurality of camera image information based on static pedestrian distribution information, the present invention is based on dynamic pedestrian distribution information (time change information) of one camera. Is fundamentally different.

  The second is an invention (Non-Patent Document 3) called “Person Counting System (IBS Counter)” by CEE Day System Co., Ltd. The technical contents are as follows. We are developing a technology that accurately identifies moving objects by recognizing optical flow estimation and extracting 3D space object movement information from the velocity field information without using the spatio-temporal difference information of the image as the subject for recognition of moving objects. . By applying this technology, we sell buildings near entrances and entrances, large halls, public customer facilities, convenience stores, elevator entry / exit counting devices and intruder detection devices. While this existing technique is a technique for accurately counting the number of people by paying attention to each person (a technique for extracting object movement information in a three-dimensional space and accurately identifying a moving object), the present invention is used as a crowd. It is fundamentally different in that it is a technology for determining the necessary conditions, focusing on the characterization of the nature of the technology, not pursuing strictness, simplifying it as much as possible without impairing its practicality.

  The third is “Comprehensive Evacuation Guidance Simulation System Study for Dense Space” by the National Institute for Disaster Prevention Science and Technology. The technical contents are as follows. As an in-facility real-time management system, the following four elemental technologies (existing people counting system, crowd simulation system, database system, guidance information display system) are pending patent application (Non-Patent Document 4). Whereas this existing technology predicts the behavior of the crowd based on the existing number of people information (static number information) in the facility that accurately measures each person using the number of people counter, the present invention is the video information of the monitor camera Is fundamentally different in that the human density on the entire screen is approximately calculated and dynamic pedestrian distribution information (time change information) is given.

Therefore, the purpose of the present invention is not to measure the number of people by paying attention to individual human beings, but to focus on characterizing the properties as a crowd, pursuing strictness, and as simple as possible without impairing practicality. It is intended to provide a technology that contributes to crowd guarding by determining the necessary conditions, grasping the congestion situation in real time.
http: // chihara. aist-nara. ac. jp / public / thesis / master / 2002 / maeda0051096. pdf http: // www. ced. co. jp / riv3. htm http: // www. pref. Miyagi. jp / hk-tihouken / sodan / jirei7-5. html, http: // www. kedm. bosai. go. jp / japanes / kenkykaihatsu / 005_nisshukukan. html

領域の面積から求める。実際の人数は画像に写っている人の数を目視により数え、画像処理から得られる人数とのカリブレーション（更正曲線）のためのリファランス値として用いた。人が重なっている場合は頭部の数を数え、対象領域の面積は、画像内の物体から縮尺の基準を定めて推定した。群集の移動方向検出はテンプレートマッチングを利用した。固定カメラにより撮影した隣接時刻のサンプリング画像のうち、新しい画像内の局所部分（位置ｓ_ｘ，ｓ_ｙサイズｍ×ｎ）をテンプレートに定め、古い画像内で上下左右に移動し不一致度を計算した。古い画像内で不一致度の低い位置から、新しい画像内の局所領域の位置へ物体が移動したと推定できる。古い画像内の位置を（ｉ，ｊ）、移動量をａとすれば、不一致度行列Ｒをａ×ａの行列として次式で計算した。

を提供する。
南敏，中村納：画像工学，コロナ社（１９８９） In the present invention, in order to solve the above-described points, the proportion of the area of the crowd in the image obtained by photographing the congestion situation at the site is used as the feature amount of the crowd density. Subsequently, a binarization operation for separating the crowd and the background is performed on the image, and a discriminant analysis method (Non-patent Document 5) is used as the binarization method. In this method, when the density t is divided into two groups, the density t _max that maximizes the variance between the groups is used as a threshold value. The purpose of this operation is to estimate the density that separates the pixel group representing the background and the pixel group representing the crowd. Since it does not operate properly when used for the entire image, the threshold is estimated by limiting the region, and this is used as the threshold for the entire image. The image is apparently distorted as shown in FIG. 1 according to the angle θ taken. When P is the camera position, OA perpendicular to the ground is AA 'and horizontal OB is BB'. Therefore, the area of the crowd is shown as cos θ times and the area of the entire image as sin θ times. Therefore, the number of pixels in the entire image is m,

Obtained from the area of the region. The actual number of people in the image was visually counted and used as a reference value for calibration (correction curve) with the number of people obtained from image processing. When people overlap, the number of heads was counted, and the area of the target area was estimated by setting a scale reference from the object in the image. Crowd movement direction detection uses template matching. A local part (position s _x , _sy size m × n) in a new image is defined as a template among sampling images taken at a fixed time taken by a fixed camera, and the degree of inconsistency is calculated by moving up and down, left and right in the old image. . It can be estimated that the object has moved from the position with a low degree of mismatch in the old image to the position of the local region in the new image. Assuming that the position in the old image is (i, j) and the movement amount is a, the mismatch degree matrix R is calculated as the a × a matrix by the following equation.

I will provide a.
Satoshi Minami, Nana Nakamura: Image Engineering, Corona (1989)

  The present invention has the following effects. From the results of the embodiment in which the crowd density was estimated by changing the place and the situation, there is a strong correlation between the actual crowd density and the feature quantity of the crowd density for each place, and in order to grasp the increase and decrease of the crowd, the present invention This method is sufficient, and can contribute to effective placement and immediate response of personnel in crowd guards. If the scope of application is determined, it can be considered as a practical element technology and can be constructed as a system.

Hereinafter, the present invention will be described in detail based on examples.
<Example 1>

On November 4, 2003, the weather was fine, and the crossing part of the scramble crossing in front of Hachioji Station North Exit was photographed for about 10 minutes from 7:00 pm (Hachioji Station Crossing (Night)). A home video camera was fixed on a tripod on the pedestrian bridge in front of the station overlooking the intersection. The angle overlooking the crowd was approximately 10 °. Next, a still image was taken out from the video image every second using a personal computer equipped with a video capture board. The actual crowd density was estimated from the area of the specific area in this image and the number of people in the area. The area of the specific region was estimated by setting an object in the image as a reference for the scale. At this time, as a result of correcting distortion in the depth direction caused by the shooting angle, the area was 149.7 m ² . The number of people in the area was counted visually. When people were overlapping, the number of heads was counted. As a result, the minimum number was 13 and the maximum number was 37. It is estimated that the crowd density changed from 0.09 person / m ² to 0.25 person / m ² . Next, paying attention to the difference in brightness between the human and the background in the image, binarization processing is performed to separate the two, and the area ratio between the person area occupied in the specific area and the entire specific area is expressed as the crowd density The feature amount. Since the specific area and the person area are distorted by the shooting angle, the feature amount obtained from the two is corrected. As a result, as shown in FIG. 2, there was a strong correlation between the obtained feature value (ratio) of the crowd density and the actual crowd density (number of people / m ² ) having a correlation coefficient of 0.96. Next, for the same shooting area, still images at different times are extracted from the video image, and a local portion is defined in the new image. The same part in the old image is defined as the initial corresponding point. For the corresponding points of both images, the sum of the square of the density difference and all the corresponding points is defined as the degree of inconsistency. The corresponding points in the old image were moved 10 dots up, down, left, and right to obtain a mismatch degree matrix. As a result, as shown in FIG. 3, the value of each element of the disagreement degree matrix is the smallest among the elements indicating the movement source of the object, and the movement direction of the crowd can be estimated.
<Example 2>

On November 5, 2003, the weather was clear and the crossing part of the scramble crossing in front of the Hachioji station north exit was photographed for about 10 minutes from 9:00 am (Hachioji station crossing (morning)). A home video camera was fixed on a tripod on the pedestrian bridge in front of the station overlooking the intersection. The angle overlooking the crowd was approximately 10 °. Next, a still image was taken out from the video image every second using a personal computer equipped with a video capture board. The actual crowd density was estimated from the area of the specific area in this image and the number of people in the area. The area of the specific region was estimated by setting an object in the image as a reference for the scale. At this time, as a result of correcting distortion in the depth direction caused by the shooting angle, the area was 149.7 m ² . The number of people in the area was counted visually. When people were overlapping, the number of heads was counted. As a result, the minimum number was 5, and the maximum number was 25. The crowd density is estimated to have changed from 0.03 person / m ² to 0.17 person / m ² . Next, paying attention to the difference in brightness between the human and the background in the image, binarization processing is performed to separate the two, and the area ratio between the person area occupied in the specific area and the entire specific area is expressed as the crowd density The feature amount. Since the specific area and the person area are distorted by the shooting angle, the feature amount obtained from the two is corrected. As a result, as shown in FIG. 4, the obtained feature value of the crowd density and the actual crowd density had a strong correlation with a correlation coefficient of 0.92. Next, paying attention to the fact that all the moving objects in the image are humans, the area ratio between the moving object area in the specific area and the entire specific area is used as the feature amount of the crowd density. The moving object is detected by binarizing density changes occurring in the same imaging region. That is, using an image at a different time, a portion where the density change of the corresponding point has increased by a certain level is detected as a moving object. As a result, as shown in FIG. 5, the obtained feature value of the crowd density and the actual crowd density had a strong correlation with a correlation coefficient of 0.84.
<Example 3>

On November 5, 2003, the weather was clear, and a passage with escalators and stairs inside the Hachioji station was photographed for approximately 10 minutes from 9:30 am (up and down the escalator inside the Hachioji station). The photo was taken with a home video camera fixed in a place overlooking the aisle. The angle overlooking the crowd was approximately 30 °. Next, a still image was taken out from the video image every second using a personal computer equipped with a video capture board. The actual crowd density was estimated from the area of the specific area in this image and the number of people in the area. The area of the specific region was estimated by setting an object in the image as a reference for the scale. At that time, as a result of correcting the distortion in the depth direction caused by the photographing angle, the area was 16 m ² . The number of people in the area was counted visually. When people were overlapping, the number of heads was counted. As a result, the minimum number was 2, and the maximum number was 9. It is estimated that the crowd density changed from 0.13 person / m ² to 0.56 person / m ² . Next, paying attention to the difference in brightness between the human and the background in the image, binarization processing is performed to separate the two, and the area ratio between the person area occupied in the specific area and the entire specific area is expressed as the crowd density The feature amount. Since the specific area and the person area are distorted by the shooting angle, the feature amount obtained from the two is corrected. As a result, as shown in FIG. 6, there was a strong correlation (correlation coefficient 0.8) between the obtained feature value of the crowd density and the actual crowd density. Next, for the same shooting area, still images at different times are extracted from the video image, and a local portion is defined in the new image. The same part in the old image is defined as the initial corresponding point. For the corresponding points of both images, the sum of the square of the density difference and all the corresponding points is defined as the degree of inconsistency. The corresponding points in the old image were moved 10 dots up, down, left, and right to obtain a mismatch degree matrix. As a result, as shown in FIG. 7, the value of each element of the disagreement degree matrix is the smallest in the element indicating the movement source of the object, and the movement direction of the crowd can be estimated.
<Example 4>

On November 5, 2003, at Hachioji Minamino Station, after getting off the train, the crowds heading for the ticket gates were fixed at the top of the stairs connecting the ticket gates and the platform, and were photographed for about 10 minutes from 10:00 am (Hachioji Minamino) Station platform). The angle overlooking the crowd was approximately 50 °. Next, a still image was taken out from the video image every second using a personal computer equipped with a video capture board. The actual crowd density was estimated from the area of the specific area in this image and the number of people in the area. The area of the specific region was estimated by setting an object in the image as a reference for the scale. At this time, as a result of correcting distortion in the depth direction caused by the shooting angle, the area was 8.9 m ² . The number of people in the area was counted visually. When people were overlapping, the number of heads was counted. As a result, the minimum number was 1, and the maximum number was 13. The crowd density is estimated to have changed from 0.11 person / m ² to 1.46 person / m ² . Next, paying attention to the difference in brightness between the human and the background in the image, binarization processing is performed to separate the two, and the area ratio between the person area occupied in the specific area and the entire specific area is expressed as the crowd density The feature amount. Since the specific area and the person area are distorted by the shooting angle, the feature amount obtained from the two is corrected. As a result, as shown in FIG. 8, there was a strong correlation (correlation coefficient: 0.94) between the obtained feature value of the crowd density and the actual crowd density. Next, for the same shooting area, still images at different times are extracted from the video image, and a local portion is defined in the new image. The same part in the old image is defined as the initial corresponding point. For the corresponding points of both images, the sum of the square of the density difference and all the corresponding points is defined as the degree of inconsistency. The corresponding points in the old image were moved 10 dots up, down, left, and right to obtain a mismatch degree matrix. As a result, as shown in FIG. 9, the value of each element of the disagreement degree matrix is the smallest among the elements indicating the movement source of the object, and the movement direction of the crowd can be estimated.
<Example 5>

On December 22, 2003, the weather was clear and the communication passage between the Keio Line and JR Line in Shibuya Station was photographed for 10 minutes from 10:00 am (Shibuya Station (connection passage between Keio Line and JR Line)). The video camera for home use was fixed with a tripod at a position overlooking the communication passage (near the entrance to the restaurant street). The angle overlooking the crowd is approximately 35 °. Next, a still image was taken out from the video image every second using a personal computer equipped with a video capture board. The actual crowd density is estimated from the area of the specific region in the image and the number of people in the region. The area of the specific region was estimated by setting an object in the image as a reference for the scale. At that time, the distortion in the depth direction caused by the shooting angle was corrected. As a result, the area was 60 m ² . The number of people in the area was counted visually. When people were overlapping, the number of heads was counted. As a result, the minimum number was 2, and the maximum number was 13. The crowd density is estimated to have changed from 0.03 person / m ² to 0.22 / m ² . Next, paying attention to the difference in brightness between the human and the background in the image, binarization processing is performed to separate the two, and the area ratio between the person area occupied in the specific area and the entire specific area is expressed as the crowd density The feature amount. Since the specific area and the person area are distorted by the shooting angle, the feature amount obtained from the two is corrected. As a result, as shown in FIG. 10, there was a strong correlation (correlation coefficient of 0.96) between the obtained feature value of the crowd density and the actual crowd density. Next, paying attention to the fact that all the moving objects in the image are humans, the area ratio between the moving object area in the specific area and the entire specific area is used as the feature amount of the crowd density. The moving object is detected by binarizing density changes occurring in the same imaging region. That is, using an image at a different time, a portion where the density change of the corresponding point has increased by a certain level is detected as a moving object. As a result, as shown in FIG. 11, there was a strong correlation between the obtained feature value of the crowd density and the actual crowd density having a correlation coefficient of 0.95.
<Example 6>

The weather was fine on December 22, 2003. The crossing part of the Shibuya station intersection was photographed for 10 minutes from 11:00 am (Shibuya station intersection). A video camera for home use was fixed on a tripod in a coffee shop overlooking the intersection. The angle overlooking the crowd was approximately 20 °. Next, a still image was taken out from the video image every second using a personal computer equipped with a video capture board. The actual crowd density was estimated from the area of the specific area in this image and the number of people in the area. The area of the specific region was estimated by setting an object in the image as a reference for the scale. At this time, as a result of correcting the distortion in the depth direction caused by the photographing angle, the area was 180 m ² . The number of people in the area was counted visually. When people were overlapping, the number of heads was counted. As a result, the minimum number was 7, and the maximum number was 75. The crowd density is estimated to have changed from 0.04 person / m ² to 0.42 person / m ² . Next, paying attention to the difference in brightness between the human and the background in the image, binarization processing is performed to separate the two, and the area ratio between the person area occupied in the specific area and the entire specific area is expressed as the crowd density The feature amount. Since the specific area and the person area are distorted by the shooting angle, the feature amount obtained from the two is corrected. As a result, as shown in FIG. 12, there was a strong correlation (correlation coefficient: 0.93) between the feature amount of the obtained crowd density and the actual crowd density. Next, paying attention to the fact that all the moving objects in the image are humans, the area ratio between the moving object area in the specific area and the entire specific area is used as the feature amount of the crowd density. The moving object is detected by binarizing density changes occurring in the same imaging region. That is, using an image at a different time, a portion where the density change of the corresponding point has increased by a certain level is detected as a moving object. As a result, as shown in FIG. 13, there was a strong correlation (correlation coefficient of 0.96) between the obtained feature value of the crowd density and the actual crowd density.
<Application example 1>

The dynamic crowd density real-time estimation method of the present invention connects an admission ticket vending machine and a dynamic crowd density estimation system with a communication function in order to prevent crowd accidents at an event venue, and excessively crowds going to the event venue. Applicable to adjust so as not to get crowded. For the purpose of this verification, we considered the student cafeteria as the event venue and examined the following system configuration with the ticket vending machine as the admission ticket vending machine. The timing of the verification experiment was selected when there were events such as academic conferences on campus. In general, visitors from off-campus use the cafeteria and sell their tickets with automatic ticket machines, but they are usually crowded and queued at noon. Therefore, a special corner is set up for visitors from outside the university. Even in normal crowds, the people in the kitchen are very busy and are full of food as they provide food to the people who come one after another. There is no person in charge of monitoring the crowd in the cafeteria. It is necessary to restrict the entrance of insiders in order to have outsiders feel comfortable eating. The change of congestion was monitored from a surveillance camera installed in the cafeteria, and the number of ticket vending machines operated was controlled according to the congestion. FIG. 14 shows the basic configuration of this system. This system comprises three food vending machines, one camera, a personal computer for data processing and control, and a communication cable for connecting them. A personal computer is equipped with a video capture board for connection with a video camera and a control board for connection with a ticket vending machine. 15 and 16 show the installation status of this system. The cafeteria is on the 3rd floor, and a ticket vending machine is installed up the stairs. A user purchases a meal ticket, enters the dining room from the entrance on the right side of the ticket vending machine, and when the user finishes the meal, the user exits from the exit. The flow of people is indicated by arrows in the figure. A bird's-eye view was shot with a camera installed in an indoor high place. The camera was controlled from the data processing device to determine the shooting range. The angle of view of the camera and the adjustment of the shooting angle were adjusted so that the entire indoor area was taken. Angle theta _c at this time shows the central portion of the imaging range. Since the angles at the top, middle, and bottom of the shooting range are different, the angle was measured by temporarily moving the upper and lower targets of the shooting range to the center of the video. Let them be θu and θd, respectively. These angles are used to correct distortion due to the shooting angle. The estimation of the shooting area was based on the scale on the upper, middle, and lower parts of the shooting range as the distance on the floor where the height of the person reflected in each part was projected. At this time, correction was performed using the upper, middle, and lower angles measured in the selection of the shooting range. The video camera was controlled from the data processing device, and the indoor scene was started. The adjustment of the image (color, brightness of the image) was left to the auto function of the camera, and the video image was stored in the data processing device through a video capture board attached to the data processing device. In the data processing apparatus, the feature amount is obtained by binarizing the brightness of the still image taken out from the video every second. Each ticket vending machine was independently controlled to either “on sale” or “cancellation of sale” according to the feature amount. The feature quantity and the number of operating ticket machines are as follows. 3 units if feature quantity <0.6, 2 units if feature quantity <0.8, 1 unit if 0.8 <feature quantity <0.9, 0.9 <feature quantity In the case of 0, 0 ticket machines are operated.

As a result, the crowd from the 12:00 to 13:00 lunch at the cafeteria on the 3rd floor of the Tokyo Tech welfare building, which was the subject of the experiment, disappeared despite the event that the conference was held, and the conference was held in the cafeteria. It was found that the confusion between related outsiders and students can be avoided, and this system is extremely effective in preventing accidents caused by crowded crowds at the event venue.
<Application example 2>

  The real-time estimation method of the dynamic crowd density of the present invention connects the dynamic crowd density estimation system and the central monitoring center with a communication function in order to prevent a crowd accident toward the event venue. Applicable to adjust at the central monitoring center so as not to be overcrowded. For the purpose of verification, Tokyo Institute of Technology is regarded as the event venue, the route to Hachioji Minamino Station, the nearest station, and the bus to the Tokyo Institute of Technology bus stop (Figure 17) to the event venue such as fireworks display As a crowd accident prevention system shown in FIG. 19, a monitor camera is placed at the position shown in FIG. 18 to monitor and guide the movement of visitors, such as a crossover or a stairway on the route. A system configuration like this was studied. At the university's exclusive bus stop in front of Hachioji Minamino Station, when crowded in the morning, there is a line that spans approximately 500 meters. Currently, several organizers are assigned to deal with the congestion. During the day, crowds before bus stops are controlled by several organizers. Buses are operated at intervals of 2 to 5 minutes during times when there are many users. In other time zones, there may be cases where only one car is coming even if it is a serious procession. Although it is possible to increase or decrease the number of buses according to the report of the bus driver, it is difficult to deal with a busy state from an empty state. The congestion situation is measured by the real-time estimation method using the image of the monitor camera of the present invention, the number of buses is increased in advance rather than in the after-response, and smooth transportation is performed from 8:30 to 11:00 in the morning. Conducted on the belt. Take a picture of the direction of the station from the ceiling of the arcade to avoid including people walking on the stairs. People walking on stairs are too close to the camera to make corrections. I took a bird's-eye view with a camera installed on the ceiling of the arcade. In order to determine the shooting range, the angle of view of the camera and the adjustment of the shooting angle were adjusted so that the area from the data processing device to just before the stairs where the camera was controlled was taken as a shooting target. The angle at this time indicates the central portion of the shooting range. Since the angles at the top, middle, and bottom of the shooting range are different, the angle was measured by temporarily moving the objects at the top and bottom of the shooting range to the middle part of the video. Let them be θu, θc, and θd, respectively. Each angle is used to correct distortion due to the shooting angle. For the estimation of the shooting area, the scale reference at the upper, middle and lower parts of the shooting range was the distance on the floor where the height of the person reflected in each part was projected. At this time, correction was performed using the upper, middle, and lower angles measured in the selection of the shooting range. The video camera was controlled from the data processing device, and shooting of the state of the matrix was started. The adjustment of the image (color, brightness of the image) was left to the auto function of the camera, and the video image was stored in the data processing device through a video capture board attached to the data processing device. In the data processing apparatus, the feature amount is obtained by binarizing the brightness of the still image taken out from the video every second. 10 to the operation instructor in the bus parked at the bus stop in the Tokyo University of Technology, which is about a few kilometers away from the image of the station direction taken from the arcade ceiling. When the average feature value per second is 0.1 or less, all LEDs are not lit, blue LED is lit when 0.1 <feature value <0.5, and yellow when 0.5 <feature value <0.8 When the LED is on and the feature quantity is 0.8 or more, the red LED lights up wirelessly, and the bus operation instructor is always equipped with one bus (portable bus operation monitoring terminal or monitoring box) This was monitored by a bus, and a departure command was issued wirelessly to the standby bus. If the feature amount is 0.1 or less, the bus operation instructor is 0 (standby), if 0.1 <feature amount <0.5 (blue LED is on), 0.5 <feature amount If <0.8 (yellow LED is lit), 2 buses, if 0.8 <feature amount (red LED is lit), 3 buses will depart from the Tokyo University of Technology bus pool and shuttle service Instructed to do. In this case, each bus will go to Hachioji Minamino Station when it receives an instruction, and when a passenger is transported to the university, it will wait in a bus pool in Tokyo University of Technology and receive the next instruction.

  As a result, the long line of several hundreds of meters lined up toward the bus stop for Tokyo University of Technology in front of the Hachioji Minamino station does not occur at all from 8:30 to 11:00. The situation of waiting for a shared bus was always visible. In addition, at times other than 8:30 to 11:00, scheduled service is performed for each predetermined season.

It is a figure for correct | amending the distortion of the image by imaging | photography angle. It is the figure which contrasted the feature amount of the crowd density obtained from the binarized image processing of brightness, and the actual crowd density in time series using an image taken at the Hachioji station square (night) landscape. It is the figure which visualized the inconsistency degree for the movement direction estimation of a crowd using the image of Hachioji station square intersection (night) landscape photography. The figure on the left is an image at a certain moment. The right figure shows the gray level where the minimum mismatch is black and the maximum is white when the position of the corresponding point is shifted up and down and left and right from the image 1/30 second before the image on the left. Coordinates correspond to the amount shifted up, down, left and right. It is the figure which contrasted the feature amount of the crowd density obtained from the binarized image processing of the brightness, and the actual crowd density in time series using an image taken at the Hachioji station square intersection (morning) landscape. It is the figure which contrasted the feature quantity of the obtained crowd density and the actual crowd density in time series based on the detection of a moving body using the image of Hachioji station square intersection (morning) scenery photography. It is the figure which contrasted the feature quantity of the crowd density obtained from the binarized image processing of brightness, and the actual crowd density in time series using the image of the escalator up-and-down landscape photography in Hachioji Station. It is the figure which visualized the inconsistency degree for the movement direction estimation of a crowd using the picture of the escalator up-and-down landscape photography in Hachioji Station premises. The figure on the left is an image at a certain moment. The right figure shows the gray level where the minimum mismatch is black and the maximum is white when the position of the corresponding point is shifted up and down and left and right from the image 1/30 second before the image on the left. Coordinates correspond to the amount shifted up, down, left and right. It is the figure which contrasted the feature amount of the crowd density obtained from the binarized image processing of brightness, and the actual crowd density in time series using the image of the Hachioji Minamino Station home landscape photographed. It is the figure which visualized inconsistency degree for the movement direction estimation of a crowd using the image of Hachioji Minamino Station home landscape photography. The figure on the left is an image at a certain moment. The right figure shows the gray level where the minimum mismatch is black and the maximum is white when the position of the corresponding point is shifted up and down and left and right from the image 1/30 second before the image on the left. Coordinates correspond to the amount shifted up, down, left and right. Shibuya Station (communication passage between Keio Line and JR Line) Using the image of landscape shooting, the feature value of crowd density obtained from binarized image processing of brightness and the actual crowd density compared with time series is there. Based on the detection of moving objects using Shibuya station yard (communication path between Keio Line and JR line), it is a figure comparing the obtained crowd density features and the actual crowd density in time series. . It is the figure which contrasted the feature quantity of the crowd density obtained from the binarized image processing of brightness, and the actual crowd density in time series using the image of Shibuya Station intersection landscape photography. It is the figure which contrasted the feature-value of the obtained crowd density and the actual crowd density in time series based on the detection of a moving body using the image of the Shibuya station square scenery photograph. It is an apparatus block diagram of the system which controls the congestion of a dining room. It is a figure which shows the correction | amendment of the distortion by imaging | photography angle. It is a figure which shows the installation condition of each apparatus in a dining room. It is a route map to the Tokyo Institute of Technology bus stop. It is arrangement | positioning sectional drawing of a camera. It is a bus operation control system block diagram.

Claims (9)

In a method for estimating dynamic crowd density in real time, a dynamic image is characterized in that a plurality of still images with temporal correlation are used, data processing for distinguishing humans and backgrounds is added to the images, and crowd density is estimated. Real-time method for estimating population density. 2. The dynamic crowd density real-time estimation method according to claim 1, wherein a still image is extracted from a video image in the still image of claim 1. 2. The method according to claim 1, wherein a still image is extracted from a video image using a personal computer equipped with a video capture board. The data processing according to claim 1, wherein binarization processing is performed on a still image extracted from a video image, and an area ratio between a moving object and the entire region in the target region is used as a feature amount of the crowd density. The dynamic crowd density real-time estimation method according to claim 1. 5. The moving object according to claim 4, wherein in the detection of the moving object according to claim 4, binarization processing is performed for density changes at corresponding points using images at different times for the same imaging region. Real-time method for estimating population density. In the feature amount of the crowd density according to claim 4, paying attention to the difference in brightness between the human and the background in the image, the binarization process for separating the two is performed, and the human area occupying the target area and the total area The dynamic crowd density real-time estimation method according to claim 4, wherein a ratio with the area is a feature amount of the crowd density. In the method for estimating the dynamic crowd density in real time according to claim 5, a still image at a different time is extracted from a video image for the same shooting target region for estimating a movement direction of the crowd, and a local portion of a new image is The position where the degree of similarity becomes high in the old image is examined, and it is estimated that the object has moved from the position where the degree of similarity is high in the old image to the position of the local region in the new image. Real-time estimation method of dynamic community density 8. The movement direction estimation of the crowd according to claim 7, wherein the index representing the similarity between two images is expressed by the reciprocal of the sum of squares of the density difference for the corresponding points of the images. The method for real-time estimation of the described dynamic crowd density. A crowd accident prevention system that controls crowding of people at an event venue, restaurant, stop, etc. using the dynamic crowd density real-time estimation method according to claim 1.

Images