Counting People in Crowds with AI
For organizers of large-scale events around the globe, ensuring the safety of visitors is of paramount importance. Understanding the movement of people at a venue is an integral part of providing reliable security and appropriate guidance to crowds. Canon has developed an AI-powered crowd counting technology that counts the number of people in an area instantaneously, which can be employed at a wide variety of locations where congestion is expected to occur.
2019/12/19 Featured Technology
Counting thousands of people in an instant
When it comes to assessing the number of people present at such crowded locations as event venues or train stations and how that number changes over time, manual counting has its limitations. In 2016, Canon released People Counter, software that uses video content analysis technology to count the number of people present in images captured by network cameras. In 2019, Canon released Crowd People Counter for Milestone XProtect Version 1.0., which not only supports the higher resolution of newer network cameras, but also boasts the ability to count thousands of people in seconds through the most recent AI technology for crowd counting.
Crowd People Counter for Milestone XProtect in action (26 sec)
During a proof-of-concept test at an international rugby match venue in 2018, Canon’s Crowd People Counter was able to count approximately 6,000 people in a few seconds. When this was compared with the number obtained by manually counting people in an image, the software successfully provided an accurate count of the crowd almost in real time with a margin of error below 5%.
A screenshot from the Crowd People Counter
Achieving high speed and precision through deep learning
In crowded places, it can be difficult for systems to obtain an accurate count of the number of people by detecting human bodies or faces, as human bodies may overlap or people may not be facing the camera. Canon’s crowd counting technology uses AI to detect and count the number of heads, making it possible to provide a precise count of the number of people in a crowd.
This crowd counting technology is just one result of Canon’s research into deep learning. At the initial stage of the study, the researchers put markings on each human head to train the AI through repetition. After that, the level of precision was improved through deep learning. This learning process involved the use of as many as several hundred thousands of sample crowd images. As it would be impossible to obtain such a vast number of actual images, computer generated 3D crowd pattern footage was created and employed to help increase the amount of data for learning.
To further enhance usability, Canon developed a lightweight deep learning model that enables processing to be performed by the CPU alone, without relying on the GPU. This contributes to lower operational costs and power consumption, which is one of the greatest advantages of the technology.
Efforts were also made to boost precision for a wide range of expected applications. For example, one of the challenges encountered was image noise under low-light conditions, which can make head detection difficult, negatively affecting counting accuracy. By working with camera hardware developers to study how noise occurs under low illumination, the developers of Crowd People Counter succeeded in detecting human heads even when there is noise in an image. The close proximity of the camera development and video content analysis development teams enabled Canon to develop a technology that realizes a high degree of precision.
By training its AI technology with image samples from a variety of angles, Canon has enabled Crowd People Counter for Milestone XProtect to detect people from images captured within an angle of depression between 10 to 65 degrees. Support for such a wide range of angle of depression allows for camera to be installed in a greater variety of locations.
Canon Crowd People Counter for Milestone XProtect is effective in both situations.
Offering network camera products with wide coverage
Capturing images of thousands of people using a single network camera requires a high resolution and a camera that can cover a large field-of-view. Jointly developed by Canon and Axis*, the AXIS Q1659 is a network camera with high image quality that realizes minimal distortion all the way to the periphery of the image and a high resolution of approximately 20 megapixels—significantly higher than that of Full HD. Using this camera in conjunction with Crowd People Counter achieves the notable advantage of enabling the counting of massive crowds across a wide area.
The Axis Q1659 network camera shown with various interchangeable EF lenses
Adding value to network cameras with AI
Crowd People Counter technology adds new value to network cameras.
For instance, by recording a time series of the number of people entering a venue, trends in crowd flow can be analyzed by time of day or day of the week. This information can be applied to marketing, where inventory can be adjusted to cater to the anticipated number of customers at a given time. It can also be employed for security purposes, to ensure the optimal allocation of security personnel. As changes in the size of the crowd can be obtained in almost real time, it also helps the relevant decision makers decide when to restrict admission to avoid overcrowding.
At the same time, it is also possible to count people within a designated area on the screen, which is useful for obtaining information on the number of people at a specific booth at an event venue or a specific area inside a station or airport.
Using crowd information for marketing applications
Canon’s people counting technology enables more effective use of images from network cameras and is now being increasingly deployed in a wide variety of fields.
Facial Recognition Technology
Recognizing human subjects under various shooting environments: image analysis technology that uses AI to identify faces
360-degree Rotatable Network Camera
360-degree, 30x zoom capabilities that hold up to constant, harsh conditions.