Facial Surveillance Deployment Complexity ExaminedAuthor: John Honovich, Published on Oct 11, 2009
Appreciating the key factors in deploying video analytics and facial surveillance are critical for setting the right expectations and achieving successful projects -- a key theme from our recent survey and recommendations on video analytic performance.
Over the past few months, we have had an ongoing discussion about 3VR's facial recongition performance based on a field test of real-time facial alerts. In this dialogue, manufacturers, integrators and consultants asked various questions and exchanged ideas on deployment issues. Most notably, the manufacturer itself posted a variety of details. I recommend reading the thread.
Summary of Facial Recognition Deployment Complexity
The following list of items are important factors discussed in our thread about making facial recognition work for real time alerting:
- Maximum width the camera can cover
- Restrictions on where cameras can be placed
- Impact of different levels of lighting
- Ensuring faces are actually detected
- Issues with obscured faces
- Changes in appearance over time
- Total cost of systems
While this discussion is not definitive, it should significantly increase the level of detail provided on facial recognition and appreciating the complexity involved.
Contrast Between Facial Searching and Facial Alerting
3VR offers both facial alerting and searching. Alerting is the use of a facial image of a person to generate a real-time notice when the person enters an area. Searching is the use of an image of a person's face to find when that person has previously been in an area (over days or weeks, etc.).
Searching is generally considered more tolerant to false matches because a searcher can quickly scan and discard the false matches. By contrast, for alerting, the monitor needs to individually review each alert whenever it arrives (usually immediately).
These considerations reviewed are extracted from the use of facial surveillance to generate real-time alerts.
Maximum width the camera can cover
While video surveillance users are accustomed to wide Field of Views (20 to 40 feet are common), the details needed for facial recognition require a much tighter field of view (FoV).
In the example cited in the Korean study, the FoV was approximately 12 feet wide (3.5 Meters) using a high definition 2MP camera. If a standard definition or analog camera was used, the maximum FoV would be half that (about 6 feet wide).
This, of course, limits what cameras can be reused given that most existing cameras are designed to capture much wider FoVs.
Also, given that most hallways and entrances with multiple doorways are greater than 6 feet wide, either multiple SD cameras or an HD camera would need to be used.
Restrictions on where cameras can be placed
The camera needs to be positioned fairly level to the faces being scanned (as shown in the sample photos).
Designers need to keep this in mind as surveillance cameras are commonly placed as significant downtilts (30 to 45 angles are common and produce top of head views). This positioning is often the easiest and least disruptive manner to deploy cameras.
Positioning cameras at angles optimized for facial recognition could incur additional cost, deployment complexity and potential aesthetic concerns.
Impact of different levels of lighting
As mentioned in the discussion, even lighting is important for real-time alerting. Even lighting produces the most consistent images improving for maximizing accuracy.
Many common issues in video surveillance can undermine even lighting. The most frequent challenge is sunlight either coming through windows or as external doors are opened. While external doorways are ideal for capturing faces as a person enters the facility (a natural chokepoint and early detection), many extneral doorways suffer from strong shadows due to sunlight surging into the building.
Ensuring faces are actually detected
3VR mentions that, "When cameras are properly and cleverly positioned, detection rates end up being ~80% for each single chokepoint camera."
To help rectify this, 3VR recommends "staging multiple choke points after another, and using 2 cameras per chokepoint."
This approach should definitely improve performance and bring detection close to 100% by increasing the chances for achieving a well lit shot with a clear view of the subject.
The most important consideration is the pure cost and complexity of deploying a series of cameras with optimized field of view, width, angle and lighting. Such a design strategy increases cost by thousands of dollars including the cost of cameras, installation and software licenses.
Accommodating Wider Areas / Entrances
Given the FoV and angle limitations, the width of areas to be covered are critical. A facility with narrow hallways and limited egresses is much simpler (and less expensive) to cover than a facility with large lobbies and numerous entrances.
As a practical example, a bank branch would be simpler to deploy than a big box retailer. Bank branches usually only have 1 or 2 entrances (though lighting could be an important issue) and a queue line. By contrast, big box retailers can have 20 - 30 foot wide entrances and very high ceilings. The combination of these two elements could require numerous cameras and special installation to obtain the minimal angles for facial surveillance.
Total cost of systems
For the Korean system, 3VR cited the cost for 4 facial recognition cameras and management as less than $50,000 USD: "1 x 3VR ServerClass IVMS and 4 3VR 2MP SmartCams, as well as the software upgrades needed, has a combined MSRP of less than $50,000"
3VR also offers an entry level appliance, the S-series, that supports up to 4 channels of facial recognition. The MSRP of the appliance with 2 channel licenses for facial recognition is $6,800 USD; cameras sold separately. Depending on traffic load and watchlist size, a more powerful appliance may be required. Also, if multiple choke points with 2 cameras per chokepoint design is used, this appliance would only cover one entry.
The cost of real-time facial alerting is certainly significantly higher than your standard IP video surveillance deployment. Now, if it catch thieves in the act, it can justify its cost. However, it does require a notable investment.
General Performance Constraints
These are important minimum elements for facial surveillance system and do not guarantee success. Performance may be impacted by attempts of individuals to avoid looking at cameras or obscure their faces. Furthermore, the quality of the facial recognition matching can vary by vendor.
These factors are important primary considerations in deploying facial recognition systems.
Related Reports on Facial Recognition
Most Recent Industry Reports
The world's leading video surveillance information source, IPVM provides the best reporting, testing and training for 10,000+ members globally. Dedicated to independent and objective information, we uniquely refuse any and all advertisements, sponsorship and consulting from manufacturers.