Increasingly, face recognition analytics are moving from servers to cameras, which have various benefits and applications. With camera hardware becoming more advanced, face recognition-enabled smart cameras are set to become more available and widespread.
Conventionally, facial recognition analytics reside on the server. But thanks to advances in camera hardware, such analytics are increasingly moving to the front end. This results in face recognition cameras, which have various benefits and applications in different vertical markets.
A main benefit of face recognition cameras is that the recognition process can be done on the frontend. “It can be widely used in fields such as police identity verification, key black and white list distribution control, facial similarity analysis, key object reminder and human trajectory analysis, saving cost for the users regarding the entire set of intelligent system deployment,” said Stanley Hu, Product Director at
Dahua Technology. “Based upon face recognition, the camera can be widely integrated with the existing business of customers, such as face-recognition-powered on-site check-in into conferences, kindergarten enrollment statistics and important visitor alerts to enhance the service experience of the customers’ existing business.”
The features and functionality included in face recognition make them suitable for safe city projects. However, there are use cases in other applications as well. “In addition to monitoring in cities, face recognition cameras can also be widely used in airport, station, building, finance, retail and other solutions. For example, in airport and station solutions, face recognition cameras can be used to distribute and control key black list personnel, arrest criminals in a timely manner and identify personnel with criminal records to make pre-judgment on malignant events,” Hu said. “In intelligent building solutions, you can deploy face recognition cameras at the entrances and exits of compounds so as to confirm the identity of the proprietors, which enables the property companies to release the barrier gates in advance, enhancing the their management experience. In financial and retail solutions, you can identify the identity of customers in advance to determine their membership identity, based upon which you can do targeted sales promotion strategy, virtually provide customers with ubiquitous service and create accurate sales opportunities.”
Hardware requirements
To run face recognition on cameras, strong computational power is required, and this had been a challenge for vendors. “The main difficulties are as follows. If you have all recognition tools on-board — that is, face capture, biometric vector modelling and comparative search — the entire facial database has to reside on the camera. But the computing resources of the camera put a limit on the file size. If you use the edge analytics only for modeling vectors, while the database is processed on the server side, then the algorithms on the camera and on the server must be absolutely identical. This will entail agreements between the camera manufacturer, the developer of the recognition algorithm and the VMS vendor,” said Alan Ataev, Global Sales Director at
AxxonSoft.
With advances in camera technology and hardware, this has become less of a problem. “In order to realize face recognition capabilities in frontends, a powerful ISP chip is required to acquire high-quality images and enable excellent image processing ability, low noise processing capability and image blur recovery. Professional AI chips need to be embedded into the cameras as well in order to realize capabilities such as face features modeling, analysis and comparison,” Hu said.
“A camera with Intel processor and GPU with at least 8GB RAM can run face recognition smoothly. This is the standard,” said Sadi Vural, President and CEO of
Ayonix. “Recently, Ayonix has developed 3D face recognition to run on Axis IP camera which has ARM-9 single core CPU in 32bit. This has been a breakthrough to run a heavy process in a conventional IP camera.”
“Since all modern algorithms are based on neural networks, it would make sense to have hardware acceleration for neural networks on the camera side. At the moment, I am aware of only one solution that can be embedded, which is the HiSilicon Kirin 970 chip with a neuromorphic processor. This is a cutting-edge device that was just released in 2017. And when they start making cameras with face recognition based on it, we can expect a noticeable step forward in this direction,” said Ataev.
Since these cameras are at their beginning stage, most end users are still taking a wait-and-see attitude. “I think the following factors may cause hesitance of customers when buying a face recognition camera: high cost, complicated parameter configuration, demanding installation process,” Hu said. “We have corresponding solutions, namely developing different series of products to meet the needs of different projects, simplifying the client configuration process, trying to make its products simple and easy to use, and delivering user guide for its customers.”