Avigilon Facial Recognition 'Appearance Alerts' Tested

By Sean Patton, Published Dec 18, 2019, 11:00am EST (Research)

******** *********** '********** ******'** *** *.*, ***** **** ****-**** facial ***********.

** *** *** **** *********, ** tested **** **** *** *** *** H5A ******* *** ****.*, *********:

  • *** **** **** ******** **** *********** work?
  • ***** **** ** ****?
  • **** *********** ***** *** **** ********* management?
  • **** ********/*** ** ******** *** **** recognition?
  • **** ** *** ******* ***** ** incidence ** ********** *****?
  • *** **** **** ** **** ** low *****/** ********?
  • *** **** **** ** ****?
  • *** **** ********** ******** ** *** Analytics ****** ** ********?

********** ****** **** ************ ******, ***** ** ****** ****.

Search ** ******

***** ********** ****** ****** *** ***** various *************** **** *****, ***, ******, age, ******, *******, ***., ********** ****** only ****** ** ******* *****. **** can ** ********* ***** **** **** similar ******.

Notice: ***** ****** **** ******

***** *** ******* **** ******, ** want ** ********* **** *** **** on * ***** **** ** ~** faces, *** **** **** *** ****** test ******** **** * ***** ****** of ****. ***** ******* ****** *********** issues, ** ********* ***** ***** ***** to ******* ******, ****** ***** ** verify *** ********* ******.

Detailed ********** ***** ********

*** ******* ********** ******* *********:

  • **** *********** ********** ** ~*****:** ***** **** **** *********** ****** reliably ** ~***** ** ****-*** ******. This ** ***** **** **** ***** systems **** ******* ****** ** ******, but ******* ** **** ** ***** with*********.
  • ****** **** ***** ***********:** *** *****, ******** ********** ***** with ****** ********** ** **** ****** and **** **** (******* **** * - **' ****, *° ** **° down ****, *** * - **°). Most ******* ** ** **** ** recognize ***** ************ ** ******* ** 90°perpendicular ** *** ******, ***** ******* Anyvision (********* ****** *********** ******) *********** **** **** **** *********** ****** (Tested)*****.
  • ***** ******** **** **% **** ********:******** ********** ***** **% ******** ** a ****. **** ** *** ***** face *********** **** ** ***** ****.
  • ******** ******** ** ********** **** ****/**********:*** ****** ******* *********** (*.*. ***, sunglasses), ******** ********** ******** ******* ********** or * *** ** **** **** tilt (~**°) *** **** ***** ********** settings, ******* **** ** *** **** time ****** * **** ** *********** at ~**° ** ******.
  • *** ***** ********* ** *** ******* angle ** *********, *** ***** ********:*** ***** *********** ** *****, ******* to **** ***** *********** ** ***** systems. ** ********** ***** ** *********** down **** ****** (~**°) ***** ******* toward *** ****** *** **** ****** confidence. ** ********** ***** ** ******* walking ** * **° ***** ** to **° **** ****. ****** **** light ***********, ** ****** ********* ***** above **°*************.

*******, ***** *********** *********** *** *****, face **** ********** ** ******* *** has ******:

  • ** ******** ****** *********:***** *** **** ** ***** ** the ***** **** ***** ***** ******* by ********* ************ *******, *** ******** image ***** (*.*. ********). ******** **** they ***** ** ****** **** ********** in *** *.*, *******, **** ** a *********** ********** *** ***** ******** who **** ** ***** *** * person *** *** *** **** ** the ******** ******** ***.
  • ****** **** ********* ****:** *** *.* **** * ***** list ** *********, ***** ******* ** all ***** ***** ****** ** * single *****. **** ******** ***** **** creating **** ******* *** ******** ****** that ******* ******** ****** ** **** person, *******, *** **** ****** *** contained ** *** ****. ******** **** they **** ** ****** ******* *** multiple ***** ** *** *.*.
  • ********* ***** *** ****** *******:** **** **** ** ****** ********* named ******* *** **** ********* ******. Avigilon **** **** ***** ** ********** this ** * ****** *******.
  • *** **** ****** ****** *********:***** ******** ********** **-*** *** ****** for *** **** ***** **** ********* images, ** ***** **** *** ****** accepts **** ***** ********** ******. ************, reference ****** **** * *********** **** tilt **** ******** ******* **** ****** (e.g. *****, ****, **** *** *******) were ********.
  • ****** *********** ***** **** ******/**** *********:** ***** **** **** ***** **** beards *** *** **** ********* ******** in *** ****** ********* *** ***** as * ********* ***** ** *** Watch ****. ******** **** **** ********** and ***** ** ********* *** ****** for ********* ** ***** *** *** not **** *** ******** ******** **** beards, ****** **** **** *** **** should ** *******.
  • ** ******** *********:** *** *****, ******* ****** ** a ****** ** ******** ** ****** displayed ** ****** ******* ***** ** used ** ******* *********** **** ********** nearly ** **** ** *** ******* themselves ******** *** *****. ******** **** they *** *** ***** *** ********* for ******** ********* ******* ** ** not ********* **** *** ****** *******.

Comparative **** *** ******

********** ****** ** ********* ******** ** the ******* ****** ** ***** ************ NVRs **** ******* ****** *********** *** free *** ********** *********** ******** ** Western ******** ********* **** ********* ****** $1,000+ *** *******.

********** ****** ******** * ***-******* **** Recognition ******* ***** ** $*** **** with * ** ******* *******.

Face ***** ********** ****** *********** *** **********

*** ********* *-****** ***** ******* **** Watch **** ********** *** *********** ***** from ******** *****:

*** ********* *-****** ***** ******* *** Appearance ****** *** **** *********** ***********, examining **** *** *** *********** ******:

Face *********** ******** ***** **** ***** ***** *****

** **** *****, ******** *********** *** solid, **** **** ********** *********** ** subjects, ** *********** ****** ** **********, such ** **** ******* ** ~**° perpendicular ** *** ******:

**** **** *** *******'* **** **% covered **** * ****, ******** *** able ** ***** **** * **** level ** **********:

** ***** **** ******** *** **** to ***** ******** ******* *********** (*.*. sunglasses, ****, *****), *******, ********** ******* from **** ** ****** ** *** levels:

*** *********** *** ********** **** ***** that **** ******* *** * *********** down **** ***** ** ****:

**** ** * ******* ****** **** we ***** ** ************ ****, ***** **** ********* **** **** a *********** **** ****.

*******, **** ***** *********** *** ***** in *** ***** *** ** *********** scenes, **** * ****** ******** ** confidence **** **** ** ******, *** no ****** ****** ** ********* ***************:

Face ***** **** ************ *** ********** ******

***** *** ** ***** ******** **** recorded *****, ***** *** *** ** the ******* *********** ** ***** ** our******** ********** ****** ****. ********** ***** *** ** ****** to ******** ****** ***** **** *** analytics *** *** ************* ******:

*** ****** ******** ******* ** ***************, and ********* *** ********* ** ****** we ***** ** *** ******** ****:

*******, **** *** ** ********* *** users ** ***** ** ** *** to ******** ****** ***** ***** ***/*** not *****, ***** ** ** ********** of ******* * **** ** ** is *** ********, *** ***** ****** be ***** *** *****. **** **** is******** ** ** ********** ********* ** the *******'* **********, ******* ** ******** color, ******/***** *****, ***.

************* ******** ***** *** **** ** added ** *** **** ***** **** from ********** ******, ** *** ********** offered.

* ********* ***** **** ***** ******** significant ***** *********, *** **** ***** list ******* ****** **** ******* **** details, **** *********** **** **** ******:

************, ** ***** **** ******** **** beards **** **** ****** ** ** rejected, ************ ** ***** ***** *** not ******* ** ********, **** ** the ***** *** **** **** ******* that ***-******* ********:

******** ***** **** *** ***** ********* image ****** **** ~******, *** **** eyes *** **** *******, *** **** requirements **** ** ********* ** *** 7.8 **** ******** ******:

********* *** ********* ***** *** ********** criteria **** *** *** **********

  • ******* ** (******* **) ****** ****** the **** ***-**-*** *** ****-**-*** ** the ********
  • ***** **** ****** ****** *** ** occluded, ****** ******* *********, **** ***** above *** *** ** *** ****, and ****** **** **** *** **** visible
  • *** **********, ***** ** * ********** and *******-**** ***** ****. ** ******** to **** **** *** ********* ** this *****. *** ****.*, ******* ***** are ***** **** ******** ***** *** watchlist ***** ******** *** *******. ** ACC *.*, ******** ********* ****** **** be ********* ** * ****** ******* quality *** *** ** **** ******* a ***** ***** ******* *****.

******* ********** **********, *** **** ***** List ****** ********* ***** *** ****** to ** ******* ** *** ***** List:

******** **** **** ** *** *.* multiple ****** ** *** **** **** and **** **** ** **********.

Duplicate ****** *** ** ****** **** ****** *******

**** *** ****** ***** **** *** duplicate ***** *********, ** ***** **** multiple ******* ***** ** ******* ** a ****** ********** ** * *******:

**** ***** ** ** ***** ** a ****** ******* *** *+ ********* images ** *** ***** ****, ***** would **** *** ****** ****** ****** list ** *** ***** ** ********* interface **** * ****** **********.

Face ***** **** ****** ******* / ** ****** *** *** **** ***********

**** ***** **** ***** ********* ** the ******** *** ***** ****** ***** well, ******* ********* *** ******* *****:

*** ******* ********** **** ** *** results *** *** ** ******** *** a ****** **** *****, ** * system **** ***** ** ******** *** day/hour **** ** ****-********* ** ****** through.

************, ***** ** ** ******** ** the **** ***** **** ********** ****** to ******* *** *********** ** * subject. ******** **** **** **** ** added ** *** *.*:

**** ******** ** * ****-**** ********** function ** **** ***** ** ****. In *** *.* *** **** ** able ** ****** ********** ******** **** the **** ********* *** **** ****** images ** ****** ********** ********.

Comments (23)

Undisclosed Manufacturer #1

12/19/19 12:52am

Since this is inside the ACC 7.4 you don’t need to stand up another server.

Agree: 1
Disagree
Informative: 2
Unhelpful
Funny
Avatar

Sean Patton

IPVM | In reply to Undisclosed Manufacturer #1 | 12/19/19 02:33pm

That is correct, you need to have Appearance Search running with NVidia GPU, and purchase the Face Recognition licenses.

Agree
Disagree
Informative
Unhelpful
Funny

Michael Miller

In reply to Sean Patton | 12/19/19 06:43pm

Slight correction. H5A/H4A, server with GPU and Face Rec License. Face Rec is not an option yet on our AI Appliance.

Agree
Disagree
Informative
Unhelpful
Funny
Avatar

Sean Patton

IPVM | In reply to Michael Miller | 12/19/19 06:49pm

Thanks Mike!

Agree
Disagree
Informative
Unhelpful
Funny
Avatar

Morten Tor Nielsen

prescienta.com
12/19/19 07:45am

Rob seems to have those airpods fused to his ears :)

I gotta say: The system performs almost miraculously well on pictures where a human would never be able to recognize the person in the picture. I am talking about the low light picture in the video where there are virtually no facial features visible. Considering the (deliberately) bad reference pictures it makes it even more fantastic.

It's both good and bad - consider the low-light situation, and assume that we were dealing with a real case here. If the system says "this is Rob", but you can't reasonably present a photo/snapshot that a human would agree presented Rob (it's just a washed out blur) - can you use it in court?

I can tell that it recognizes remarkably well, so did the system ever fail to recognize anyone, or does it always provide the best possible match when someone walks by? E.g. you had 30 people in the reference database, and 70 that were not. When one of the 70 walks by, does it say "Oh, here's someone we don't know" and would that ever happen (mistakenly) if one of the 30 walks by in low light, recently shaved, new hair-cut etc. Basically how often does it miss a detection?

Is there any training option? If the system says "here's Rob" but in reality it's Mike. Can you then tell the system "NO! Bad system. This is NOT Rob!"?

Agree: 4
Disagree
Informative
Unhelpful
Funny: 1
Avatar

Sean Patton

IPVM | In reply to Morten Tor Nielsen | 12/19/19 06:09pm

I agree that performance was very strong, especially considering (as you noted) how bad some of the reference images were, on purpose. One significant limitation (which we noted Avigilon said they would be fixing in one of the next updates) is reference images are limited to only faces capture on the connected cameras, which can definitely impact capturing high quality reference images.

I suppose one work around would be to print out an image of the subject, hold it up in front of one of your cameras and manually enter the face from that image. We did not test that, and Avigilon does not recommend that, I will see if I can get that to work today.

Regarding "missed" recognition, if a camera detects a person, and they are not matched to a face in the Face Watch List, the only the Person detected alert is displayed. If a person is detected and matched to a face in the Watch List, only the Face Watch list alert is displayed. There is no "Unknown" or "Unmatched" face alert.

Related to how often it misses a detection, that depends on a couple factors. The minimum match confidence level, and the effectiveness of the H5/H4 camera analytics (tested here), which perform very well. If you were to require a high or medium confidence level, to avoid some nuisance false positives, based on our testing you would not receive a face match alert for someone wearing a mask. However, the camera is still detecting a person in those situations, so a "Person Detected" alert is still triggered by the system. This is the case for any face capture where the minimum confidence level is not met.

There is not any training for specific faces.

Agree
Disagree
Informative: 1
Unhelpful
Funny
Avatar

Sean Patton

IPVM | In reply to Morten Tor Nielsen | 12/20/19 06:42pm

I set up a reference image of Rob holding the printed image as it one of the easiest ways to capture a clear, front-facing image of him in the system. This is not a recommended Avigilon work-around, I just wanted to see how the system would handle it.

It worked, matching Rob at a medium confidence level:

However, it matched at a high confidence level with a "real" captured image:

The PPF is lower in the printed reference image, but those are the results if you had to enter a reference image to the Watch List for someone that had not been captured on your cameras.

Agree
Disagree
Informative
Unhelpful
Funny
Avatar

Ross Vander Klok

IPVMU Certified | 12/19/19 05:14pm

Any integration with Lenel now or on the roadmap with 7.8? We have looked at Anyvision as an add on, but if our existing Avigilon can do the same thing it would be pretty sweet.

Agree
Disagree
Informative: 2
Unhelpful
Funny

Michael Miller

In reply to Ross Vander Klok | 12/19/19 07:20pm

What are you looking to so with the integration with Lenel? Also, any thoughts to replace Lenel with ACM to get all the integration your looking for.

Agree: 1
Disagree
Informative
Unhelpful
Funny
Avatar

Ross Vander Klok

IPVMU Certified | In reply to Michael Miller | 12/20/19 03:31pm

Basically using your face as your credential. "Look at the camera to unlock the door" type of deal.

Lenel is our global standard so zero interest (at this point) even looking at anything else.

Agree
Disagree
Informative
Unhelpful
Funny

Michael Miller

In reply to Ross Vander Klok | 12/20/19 04:38pm

Just make sure if your using Face Rec to unlock a door that the solution you use has liveness detection. Other wise anyone can grab you face from your Facebook account and load the pic on an IPAD and bam your doors open...

Agree
Disagree
Informative
Unhelpful
Funny
Avatar

Ross Vander Klok

IPVMU Certified | In reply to Michael Miller | 12/20/19 08:24pm

I thought Anyvision had that tackled if I recall correctly, but good point and concern! Hopefully, Avigilon figures it out as well.

Agree
Disagree
Informative
Unhelpful
Funny

Michael Miller

In reply to Ross Vander Klok | 12/20/19 08:45pm

Not with Better Tomorrow only SESAME.

Agree
Disagree
Informative: 1
Unhelpful
Funny

Undisclosed #3

In reply to Ross Vander Klok | 12/23/19 04:35pm

I was able use my iphone with a photo to open their turnstile at ISC.

Agree
Disagree
Informative: 3
Unhelpful
Funny
Avatar

Sean Patton

IPVM | In reply to Ross Vander Klok | 12/19/19 07:21pm

I emailed Avigilon about this and will post their response.

Agree
Disagree
Informative
Unhelpful
Funny
Avatar

Sean Patton

IPVM | In reply to Ross Vander Klok | 12/20/19 02:18pm

Avigilon responded, with the summary of Lenel integration:

Video Gateway with Lenel OnGuard

Release Summary

  • This is a new integration between Avigilon Control Center 6 (ACC) and later, and Lenel OnGuard 7.2, 7.3, 7.4 and 7.5.
  • It maintains the previous functionality, allowing events from the Lenel OnGuard system to be monitored and recorded through ACC alarms, as well as the ability to distinguish between individual inputs on input control models such as the LNL-1100.

Video Integration with Lenel OnGuard

Release Summary

  • This is a new integration between the ACC system and Lenel OnGuard 7.2, 7.3, 7.4 and 7.5.
  • It maintains the functionality provided by previous integrations, allowing users to access ACC features within the OnGuard interface. Users can view live and recorded video and events, export video, and control PTZ devices and monitor analytics events including bounding boxes through the OnGuard software.

Let me know if you have other specific integration feature questions.

Agree
Disagree
Informative: 1
Unhelpful
Funny
Avatar

Ross Vander Klok

IPVMU Certified | In reply to Sean Patton | 12/20/19 04:01pm

Cool, thanks for sending that on Sean! Just replied to Michael above that what I am really after is if there will be integration with Lenel that allows Avigilon face recognition to unlock doors within Lenel.

Agree
Disagree
Informative
Unhelpful
Funny

Undisclosed Manufacturer #2

In reply to Ross Vander Klok | 12/21/19 01:50am

Another option for this that's a Lenel partner is StoneLock: StoneLock StoneLock Pro - Lenel.com

Agree
Disagree
Informative: 1
Unhelpful
Funny

Undisclosed #4

02/08/20 09:46am

So does this work with Third Party Cameras or only Avigilon Cameras?

Agree
Disagree
Informative
Unhelpful
Funny

Michael Miller

In reply to Undisclosed #4 | 01/20/21 12:53pm

With the new AINVR it works on 3rd party cameras.

Agree
Disagree
Informative: 1
Unhelpful
Funny

Undisclosed Manufacturer #5

01/20/21 11:38am

Anyone knows who Avigilon is using as a face recognition partner?

Agree
Disagree
Informative
Unhelpful
Funny

John Honovich

IPVM | In reply to Undisclosed Manufacturer #5 | 01/24/21 05:38pm

Avigilon's response:

We have partnered with Paravision for our FR models. We did this because of the exceptional performance of their models for 'faces in the wild' 1:n use-cases. You can see that performance relative to a lot of the other options in the NIST 1:n results.

The FTC recently made a determination about Paravision's technology and some of the data they used for training. Paravision have been anticipating this possibility for some time and have long since taken actions to redo their machine learning models with fresh datasets and MSI/Avigilon have been using these models from the start, meaning that this FTC judgement has no impact on our product.

Agree
Disagree
Informative: 1
Unhelpful
Funny

Undisclosed Manufacturer #5

In reply to John Honovich | 01/24/21 07:18pm

Appreciated

Agree
Disagree
Informative
Unhelpful
Funny
Subscribe to IPVM Research to read the full report.
Why do I need to subscribe?
The IPVM Research Service includes products tests and shootouts plus competitive and financial analysis, helping decision-makers better evaluate purchasing, partnering, developing, and/or competing against companies in physical security.
Already have an account?
Loading Related Reports