This is the User Manual for SAFR Camera. This document describes capabilities specific to the SAFR Camera. Please refer to the SAFR Software Administration Guide for more general software capabilities that are applicable to SAFR Camera and 3^rd Party Camera Surveillance solutions as well as SAFR SCAN, the Access Control Solution.

1.2 About SAFR Camera

SAFR Camera brings NIST-leading live video face recognition and video analytics features to the edge. Embedded directly into the camera, SAFR utilizes onboard compute functionality for face detection, recognition and other analytics — eliminating the need for servers dedicated to processing video streams.

By eliminating expensive server hardware, SAFR SCAN significantly reduces the high cost and maintenance of GPU enabled server class hardware to process video and perform facial recognition. It also avoids the need for transmitting high quality video streams over the network. Finally, SAFR Camera improves reliability by bringing the watchlist and face matching to the edge.

While SAFR Camera can operate as a standalone camera with no network connection, it is also possible to take advantage of the powerful multitiered architecture of SAFR Server to provide powerful administration tools. SAFR Cameras can be managed centrally by SAFR Server or SAFR Cloud. SAFR Server or SAFR performs the following roles:

Distributed Camera configuration (single or bulk updates)
Firmware updates
Centralized watchlist management
Event aggregation for monitoring and reporting
Event to action workflows

1.3 SAFR Architecture

SAFR Cameras support either on premises SAFR Server to keep all your data on-site or SAFR Cloud for simple serverless deployments. The graphic below depicts each of these configurations.

SAFR Cloud Deployment (Serverless Deployment)	SAFR on premises Deployment (Control your data)

With a SAFR Cloud deployment, the only hardware on premises is the SAFR Camera. No PC is required. Bandwidth requirements to SAFR Cloud are limited to face recognition events and device status.

With on premises deployments, all data remains on site.

1.3.1 Hybrid Architecture

SAFR Camera can also be used in a hybrid deployment. A hybrid deployment allows for SAFR Server to share limited data with systems outside of the local network. This may be useful for connecting multiple locations or providing a means for operators to connect to SAFR Cloud for receiving notification and viewing events.

The diagram below demonstrates a hybrid deployment. SAFR Server on premises connects to SAFR Cloud which can facilitate sharing of data between networks.

A diagram of a cloud server

Description automatically generated

An operator can receive notifications, browse events and manage watchlists from a mobile device that is not connected directly to the same network as SAFR Server and the SAFR Cameras are hosted.

1.3.2 Advantages of 4k on-board processing

Most IP video cameras are 2 MP (1920x1080). For surveillance, this resolution of video offers a good tradeoff between cost and quality. Costs associated with video are due to storage size and processing power. As the video doubles in width, the storage and processing requirement typically increase by 4x. Thus, 4k video feeds, while only 2 times wider, take up 4 times the storage and processor power.

A 4k (8MP) video stream takes up to 15 Mbps of network bandwidth for each camera. A network with just 20 cameras would require 120 Gbps of network traffic. 50 cameras would saturate a 1 Gbps network. But when processing is moved to the edge, video need no longer be transmitted over the network or processed on a CPU. The SAFR Server is just there to aggregate a low bandwidth (50 kb per recognition event sent to server).

For software-based facial recognition , server class hardware is required with fast CPUs and expensive GPUs. A top of the line GPU costing thousands of dollars will process 12 to 15 4k video streams and require either Silver or Gold Xeon CPUs. Systems like this can cost $10,000 to $15,000. But with SAFR Camera, processing is kept on the edge and server is simply aggregating event data, avoiding the need for expensive GPUs and fast CPUs.

1.4 SAFR Camera vs. SAFR Software-based Facial Recognition

SAFR provides two products that perform facial recognition:

SAFR Software

SAFR works with 3^rd party video cameras (USB or RTSP) to process video in software for the purpose of performing facial detection, facial recognition and other analytics in software. This is most often done with the help of NVidia GPUs which enhance the speed and efficiently of the software solution to scale to many cameras. In addition to analytics, SAFR Software manages watchlists, aggregates events and other management functions.

SAFR Camera

SAFR performs facial detection, facial recognition and other analytics on the camera. The SAFR Software is used to manage watchlists and aggregate events but does not have the heavy task of processing video.

SAFR offers two products for performing facial recognition. Both products perform fundamentally the same actions and have a great deal of overlap. SAFR Desktop and SAFR Server Web Console provide a single interface that works with both cameras.

Both products interoperate with the SAFR Server. Each product is not mutually exclusive. A deployment can have a mix of 3^rd party cameras and SAFR Camera. There are some features that are only available in software. They are:

Person Body Detection – Detect a person’s body which is useful to measure true dwell time (i.e. if a person turns around, we still track as same person when their face was visible) and occupancy counting (counting people both as they enter and as they leave an area). Person body detection is on the roadmap for SAFR Camera.
2D Liveness – Ability to detect if a face is photo or a live person at the camera.
VMS Integrations – SAFR developed VMS integrations are done in software and not yet available with SC800 camera. SAFR Actions based integrations exist that can be made available as sample code that integrators can deploy, but these are not part of the SAFR product line as of yet.

1.5 SAFR Camera Key Features

1.5.1 Accuracy

SAFR has the fastest and smallest facial recognition amongst submitted NIST algorithms making it possible for high accuracy facial recognition on low power solid state devices.

99.87% precision in LFW faces FNMR
Accuracy with masked faces: 96.16% TPR at 0.0047% FPR on masked faces
Face match for Live Watchlist: < 100ms Average Response Time
Face search / Forensics analysis: 500,000 events/second

1.5.2 Offline operation

SAFR Camera are completely self-contained. Should the camera lose its network connection for any reason, it will continue to function using a locally cached watchlist. SAFR Camera stores the entire watchlist on the device. The cameras only need the network connection to receive watchlist updates, perform configuration or transmit events. During the time a camera may be offline, events are cached on the device and pushed to the server once network connectivity is restored. Up to 30,000 events can be cached, the first 2,000 with images.

1.5.3 Centralized Administration

SAFR Cameras are connected to SAFR Server or SAFR Cloud at installation time. Cameras can operate independently but this would mean managing the watchlists and camera configuration one by one. By connecting to a centralized system, cameras can be grouped into zones and managed centrally.

1.5.4 Bulk Camera Configuration

Settings can be applied to a single camera or to multiple cameras in bulk operations. Configuration can be applied by zones or to all cameras in a single operation. Once configured, watchlist are distributed to all cameras according to zone definitions.

1.5.5 Event Aggregation

As SAFR Camera matches each face, event information is sent to SAFR Server or SAFR Cloud where it can be centrally viewed and filtered by attributes such as site (e.g. building or other grouping location), source (camera), time, id class (threat, concern, stranger), age, gender and more. Reporting can be performed on historical event data.

1.5.6 Forensic Search (Search by Image)

A search of a face image can be performed against the event history or person database to find when and where the face appears. This is particularly useful when a face image was not already enrolled before an incident involving that person occurred.

1.5.7 Event to Action Triggers

Event can be used to trigger actions such as SMS or Email alerts. SAFR Actions is a tool that can be used to trigger any type of action such as:

Running a custom script
Make a REST API call to an external system
Play a sound or trigger a tower light
Generate a custom log of interesting events

1.5.8 Other Features

SAFR Camera also supports the following:

4k video processing (cover 2x wider area with same accuracy than 2MP cameras)
Options for 6 to 22mm or 5 to 50mm motorized varifocal lens
- High accuracy recognition up 100 feet (30m) or 200 feet (60m)
Face prioritization auto-exposure for perfect lighting to the face for best recognition results
Automated behavior with advanced control over shutter speed and aperture
Highly secure to protect personally identifiable information
- Data is encrypted at rest (on disk) with AES-256 and RSA-2048 ciphers
- Data is transmitted over TLS (https) to encrypt all transactions
- Individual user accounts controlled by an access control list (ACL)
- Device protected by secure boot that prevents rogue OS loading
Flexible Delivery: On premises, cloud & hybrid deployment models
Characterization: Anonymous age, gender, gaze, occlusion and sentiment detection
Horizontal Scalability through server clustering