How advances in edge computing are making facial recognition easier to deploy
By Ján Záborský, communications manager at Innovatrics
Recent advances in producing relatively inexpensive chips designed for AI and edge computing are transforming the way the technology is being used. This applies to biometrics, especially to facial recognition technology, which has improved significantly over the past few years. The main breakthrough came when neural networks and machine learning started to be used in the field. Both of these areas have benefited in a huge way from abundant computing power, especially in the form of GPUs which are optimized for fast floating-point math.
The traditional way of deploying facial recognition in a CCTV world was to transfer raw video streams to the server, which bore the brunt of the processing workload. It was detecting faces in each stream, extracting their features to the so-called templates, and then comparing the templates against the database for matching. The more streams (and faces), the higher the computational load. The bottlenecks of this approach are obvious: processors, even when aided by high-end GPUs, were constrained by available RAM, disk speeds and bandwidth that was needed to transfer the streams from cameras to the server.
This made scaling difficult: each new camera consumed more RAM and bandwidth, which meant necessary hardware upgrades. Transferring to the cloud — a typical approach when scalability was the issue — was not an option due to constant high bandwidth demand. However, all of these constraints fall when you can just transfer most of the processing to the edge — the cameras themselves.
Recently, Nvidia and other companies have started to produce devices and chips aimed specifically at edge computing. Devices such Nvidia Jetson are able to turn standard CCTV cameras and other devices “smart” simply by plugging them in. This makes upgrading an existing CCTV setup a much more affordable option, as you don’t need to replace all the cameras, but instead, just connect them to a specialized edge device.
Even the first-generation edge devices helped considerably with computational load; they were able to detect faces and extract their templates on the device, transferring just the templates to the server for matching. Detection and extraction count for 80% of the workload, while also getting rid of the need to send the whole stream for processing to the server. This alone meant a drop in bandwidth requirement by over 90 percent.
Both effects taken together then effectively meant that real-time facial recognition in the cloud has become possible and that the CCTV systems, if facial recognition was necessary, have become scalable basically without limits.
The current edge devices are perfectly capable of running real-time facial recognition directly on the camera, including matching. To get an idea about the available speeds, a camera outfitted with an Ambarella CV22 chip is able to fully process a face in the video stream in about 70 ms. Most of that time is taken by detection (approximately 22 ms) and then template extraction (less than 20 ms). Of course, these steps need to be done for each face in the stream. These times are comparable to the operations being performed on a server.
With properly optimized algorithms, the matching itself on the camera side is almost negligible: it takes 2ms to identify a face in a gallery of five thousand and less than 10 ms in a database of over 30 thousand faces. The other chips such as Blaize Pathfinder or Nvidia Jetson Xavier NX show similar results.
Such results allow for new use cases for facial recognition such as easy access control solutions in smart buildings or efficient security solutions for high-security spaces such as airports or border crossings. Moreover, it makes the possible approaches much more modular. Depending on the use case, the whole video stream can be discarded with just faces extracted and matched locally and only a relevant message sent to the server (e.g. “Authorized person found and let in”). This can address one of the biggest concerns when using biometric systems — sensitive data protection, storing and vulnerability to their leaks.
This also makes cloud deployment of real-time video facial recognition a viable proposal, harnessing the cloud’s no-maintenance and high scalability capabilities. Moreover, it makes the security aspect of biometric authentication more accessible to small and medium enterprises or fast-growing startups. Biometric authentication avoids many security pitfalls from easy-to-guess passwords or PINs to buddy-punching with access cards. In fact, passwordless authentication is becoming a preferred method of accessing computers as well and is being spearheaded by companies such as Google or Microsoft. Biometric-based building access authentication — among other use cases — fits right into this ecosystem, which is poised to become an accepted standard.
About the author
Roman Ševec is the marketing manager at Innovatrics, a multimodal biometrics solutions provider based in Slovakia.
DISCLAIMER: Guest posts are submitted content. The views expressed in this post are that of the author, and don’t necessarily reflect the views of Edge Industry Review (EdgeIR.com).
Article Topics
Ambarella | Blaize | edge AI | facial recognition | Innovatrics | Nvidia
Comments