SAFR has licensed their technology to edge device manufacturers and plans on bringing AI to the edge with SAFR running on the SOC.
In the post-
COVID-19 world, the ability to manage crowds will gain greater prominence on the agenda of every head of security and operations. The need to resume operations safely and quickly will be directly linked to how well they can manage crowds.
“The market will be shaped by technologies that give visibility to what is happening on the premises. Video feeds will be useful information and can give additional data for the security manager,” explains Daniel Grimm, GM of
SAFR, a Realnetworks company that specializes in live video facial recognition.
Grimm is not a typical security industry executive. With a background in business consulting and product management for Amazon, he is a relative new comer to the security industry. Then again, SAFR is also not a typical security industry company.
A solution looking for a problem
“We have a legacy of excellence in video,” tells Grimm. “SAFR is part of Realnetworks, which was the first company to stream audio and video online, back in the 1990s. Our facial recognition started as a feature in another consumer product, a way for users to sort photos in a photo album. The team exceeded the accuracy and performance levels set by the company CTO and then they realized they developed something extraordinary – an ability to create an algorithm superior to what the market had.” The team’s work on creating a new feature matured into a whole new product which performs exceptionally well on wild video.
High accuracy in wild images
What distinguishes SAFR from other
analytics is high accuracy in wild images (video). The company submitted its solution to the NIST (National institute of Standards and Technology) where it ranked as having the highest effective accuracy rate for live video.
“When you run on video, what is important is speed and accuracy, this means we can process multiple recognition calls of the same individual over subsequent frames in the time it takes other algorithms to complete just one. This delivers an accurate result in real time (less than 150 ms). There is no limit to the amount of people in the image, the only constraint is the amount of compute power available. Working with NVIDIA GPUs, we can process over 20 video feeds on one tesla core, which can process about 10 people appearing simultaneously in the frame” explains Grimm. In 2019 SAFR was rolled out in India in less than 24 hours to help secure the area where PM Modi was about to give a speech and check people’s faces against a watchlist of persons of concern. Luckily no one triggered an alert during the event. “The larger the database, the more likely you are to have a false positive. We tested SAFR against lists containing millions of faces, but in practice most of the use cases require a much smaller watch list running up to a few thousands” added Grim.
Product positioning
SAFR is not just an algorithm, but rather a full stack video analytics solution that can be a standalone application or integrated with an existing VMS. Currently most of the installations are an integration with an existing VMS and not a stand-alone application. In terms of deployment, it can be cloud-based, on premises or a hybrid cloud and on-prem solution.
SAFR is targeting the verticals in the physical security space where the need for high accuracy real-time
facial recognition for video is critical for end users, namely law enforcement, critical infrastructure, gaming and airports.
Future roadmap
In its next phase, the company is working on improving management of the watch list and secure access. “In the product today, we have facial recognition, age and gender classification, sentiment analysis and other features. We can associate a face to a body and tag it in a way that allows tracking even if face is not visible. Going forward, we want to build on our strengths, the low TCO and ability to be embedded on edge devices. We have licensed our technology to edge device manufacturers and plan on bringing AI to the edge with SAFR running on the SOC,” concluded Grimm.