The world's first Crowd Behavior Analysis Technology that detects congestion status and anomalies from crowd footage!
- MiyanoFor example, if you saw footage showing a large number of people and were asked about how many were there, I believe you'd answer with an estimate rather than counting each person. We do the same thing, only with computers.
- MitaCan you be more specific?
- MiyanoFirst, take a look at this slide (he opens up a slide on the computer).
- MitaIs this a still image from a security camera?
- MiyanoWe hired extras to shoot this, but imagine it is an image taken from a security camera. In a photo like this that shows many people, we organize them into rough groups in the following way (shows the next slide).
- MitaIt looks like a grid of small green squares has been applied to the same security camera image.
- MiyanoWe analyze each of the groups divided by green lines as a cluster (shows the next slide).
- MitaOh, "Crowd images generated artificially via simulation" appeared to the right.
- MiyanoWe actually have crowd images depicting different numbers of people, such as five or three, that we created in advance. Those are the crowd images on the right. For example, if there were five people, they would overlap in a range of different patterns, right? The computer recognizes these overlapping patterns, and can recall crowd images for each number of people. When analyzing security camera footage, these crowd images and the security camera clusters are compared, and an overall picture of crowd status is produced by estimating the number of people. For example, if a cluster resembles a crowd image for five people, the computer estimates the number at five, and if one resembles a crowd image for three people, then that number is added to the estimate.
- MitaI see. I think I'm slowly starting to understand how the system works. But there must be a wide variety of ways for people to overlap, such as people that are taller or shorter, and those that are far apart or close. About how many crowd images have you created in advance?
- MiyanoWe have several hundred thousand.
- MitaWow, that's a lot!
- MiyanoWe brought over 100 extras together for the shoot, and adjusted those images automatically to increase the number of variations. Training a computer with this large number of synthetic images enables it to analyze the density and flow of people from security camera footage, making it possible to judge crowd status more accurately. It can estimate the number of people with a margin of error of about 10%.
- MitaThat's pretty accurate!
Quick detection of congestion status and anomalies from crowd footage!
- MitaWhat kinds of things does the use of Crowd Behavior Analysis Technology make possible?
- MiyanoFirst, you can accurately judge congestion status by estimating the number of people from security camera crowd footage. Using conventional analysis technology, it isn't possible to get a clear picture of congestion status. This is because conventional technology involves the computer storing a background image, and comparing it to security camera footage to find the difference. In other words, its determination is based on looking at how many openings there are between the background and people.
- MitaSo it's only a rough estimate.
- MiyanoFor that reason it can only make broad categorizations, such as "no people," "people but not crowded," and "crowded." And, of course, it can't tell how many people are there. On the other hand, as mentioned earlier, Crowd Behavior Analysis Technology can analyze crowd footage and estimate the number of people.
- MitaAnd when you know roughly how many people there are, you can tell how crowded it is!
- MiyanoThat's right. In addition to the rough number of people, you can also determine the density and flow of people, providing a more accurate picture of congestion status.
- MitaWhat other things can be done?
- MiyanoDetermining crowd status in real time enables the swift detection of anomalies in the crowd.
- MiyanoLet me show you a demonstration video to make it easier to understand. First, let's take a look at a situation where people stop or surround someone. Here goes. (The demonstration video on the PC starts)
- MitaThis looks like security camera footage. People are walking... Ah, one fell over! The other people in the area have grouped up around him, looking concerned.
- MiyanoSee how the color of the people standing around the person changed?
- MitaYes. The people grouped together are shown in red, and the value for the green graph on the left side shot up!
- MiyanoThat area is red because the flow of people stopped due to the high density there. The red line you can probably see in the green graph to the left indicates the threshold value. Crossing the threshold value indicates an anomaly, which is exactly what has happened here.
- MitaAmazing! So you can see at a glance when an anomaly occurs from the graph and changes in the color of people! That means surveillance staff can head to the scene straight away, too.
Menu starts here.
End of menu.