Avigilon Unusual Motion Detection (UMD) And Advanced Appearance Search Examined

By: IPVM Team, Published on Apr 06, 2017

Deep learning is a big trend at ISC West and Avigilon is one of the bigger names capitalizing on it.

Avigilon is showing 2 upcoming features:

  • Appearance Search capabilities that extend capabilities from just searching people to faces and vehicles
  • An "unusual motion detection" feature designed to alert users to potentially suspicious or noteworthy activity.

In this report, we examine these new capabilities, including functions and limitations, server/camera requirements, and anticipated release dates.

**** ******** ** * big ***** ** *** West *** ******** ** one ** *** ****** names ************ ** **.

******** ** ******* * upcoming ********:

  • ********** ****** ************ **** extend ************ **** **** searching ****** ** ***** and ********
  • ** "******* ****** *********" feature ******** ** ***** users ** *********** ********** or ********** ********.

** **** ******, ** examine ***** *** ************, including ********* *** ***********, server/camera ************, *** *********** release *****.

[***************]

Appearance ****** - ********

******** *** ***** * general ****** *****, ******* that ****, ***** ******, boats, *** *********** (***** other ********) *** *** still ********** ****** ** a '*******'. *******, **** selecting * ******* ** initiate ** ********** ******, the ******** ***** ********** specific ** **** *******, such ** ****/*****, *** color, **** ******* ** return ******* **** *** similar ** *** ******** object.

** * ************* ** their *** **** **** booth, ******** *** ******* an ******* ** ********* for * ****** ***, with *** ******** ********* through ******** ***** *** returning ******* ******* ** vehicles **** **** ******/***** in **********, *** *** the ***** ******* *** dimensions ** ** ***:

********* *** ****** ****** multiple ******* **************, *** the ******** **** ****** matches *** ***** **** vehicles ** * ******* appearance, ****** ** ********** level, ** *** **** the ******** ****** *** vehicle ** *** ***** matches **** *** ******** searched ***. ******* *** be ******* ******* ** clicking ** ****** **** the ******** ******** ** be *** ****** *******, which **** **** ********* the ***** ********* *******, with *** **** ***** to **** ** **** video ******** **** ******** cameras ******* *** ******* vehicle.

***** ************ *** ******** to ** ******** ** the *** *.* *******, due ** *** ****.

** ******** ** *** Appearance ****** ********* ** shown ** ********'* ********* video *** *** *******:

Appearance ****** - *****

******** ** **** ********* Appearance ****** ** ****** match *****, ****** **** is *** '****** ***********'. In *******, **** ********* for * ****** *** software **** ** **** to ****** ***** ** the ******, *** **** use *************** ** ***** faces ** ********** ****** criteria. **** ******** ********** the *********** ** ************ enroll ****** ** * face ******** *** ******* tight ****, *** **** also ****** ******** *************.

*** ******** ** *** facial ****** ********* **** rely ** *** ****** getting **** ****** ** faces, **** *** *** initial ****** *** *** any ********** ***** ***** the ******* *******.

**** ****** ** ******** to ***** ********* ** search ***** **** * longer **** ***** ** find ******* *** * person, *** ** ******* results **** ******** ****** are ******* *********. *** current ******* *********** ******* primarily ** ********, ***** reduces *** ******* ** search *** * ****** across ******** **** ** video, ** ***** ******** will **** ****** *******. It **** ****** ************* in ********* ** ***** settings ***** **** ****** may ** ******* ********* (wearing ********).

****** ****** ************ ** not **** * ****** release ******* ** **** at **** ****, *** are ******** ****** *** end ** ****, ********* to ********.

Appearance ****** ***********

********** ****** ** ********* from ******** ***** ****, meaning **** ** ******** cannot ***** ** * person ** ******* ** live ***** *** *********** begin ********* *** ***** appearances ** **** ******.

************, *** ******** **** searches ******** *****, ** an ******** ****** *** Appearance ****** ** *** a **** ****-****** ** where * ****** ** vehicle *** ** ** real ****.

Appearance ****** ************

********** ****** ** ********* in *** ********** ***** of *** ******* *+, and ******** *** **** to ******** ** ********** NVIDIA *** **** *** each ****** *** ********** will ** ***** ** (~$400).

Unusual ****** *********

********* **** ************* ** in * '*******' ****, with ** ******** ******* commitment. ******** **** ** will ** ********* ** a ******** ****** ** H4/H4SL ******* **** ********.

******* ****** ********* (***) will *** **** ********** license ***** *** *** cameras ** ***, *** will *** ******* *** hardware, ** ** ** camera-based.

*** **** ** ************ learns *** ******* ****** patterns ** * ****** over *** ****** ** a ***, **** **** results ***** ** * minimum * **** *********** period. ** **** *** rely ** *********, *** instead **** **** ****** data. ******** **** **** gives ** *** ******* to ****** ******* ******** that *** ******* ****** outside ** **** *** analytics *** ******** ** recognize/classify.

"*******" ****** ** ******* as ******** **** ******* significantly **** *** ******* baseline ** ***** ** time **** *** ******** occurred, **** ** *** activity ** *** *****, and **** ** *** objects ** ******. ** an *******, ** *** system ********* **** ****** on * ******** *** a *** **** ****** on *** ********, **** car ***** ** "*******" due ** *** **** in *** **** ** the ********. *********, ** a ****** ****** *** the ******** **** * grassy **** **** ***** be *******, *** ** activity ********* *** ***** observed ******* ** *** sidewalks.

****** ** ******* ******** are ********* ** *** cameras ** ****-****, ****** Avigilon ******* *** ******* mostly ** * *** for ********* ** ********* potentially *********** *****, ***** eliminating **** ***** ********* general ****** ********.

*** ********* ***** ***** Avigilon's ***** *** *** (note **** *** ***** callouts ***** ********** *** specific **** ** ******* motion *** *** ** the **** ***** *************):

Focus ** ***-***** *********

**** ** ***** ************ focus ** ***** ********* capabilities ** **** ** benefit ***** *** *** less ********* ***** ********** alarms *** ******, *** more ********* **** ****** their ************ ******* **** efficient *** ********* *****, or ***** ***** ** potentially *********** **********. **** should **** ******* *** appeal ** ********'* ********* capabilities, *** ******** **** users ** ***** ********** money *** *********-******* ******* vs. ***** ***** **** competitive ******* (*********/*****).

Increased *********

****** ***** ******** *** not **** ****, **** do ******* ** ********* benefit ** *** *******'* market ********* ** *** potential ****** ** ******** (e.g., ****** **** *** *** Avigilon?) ** **** ******** Avigilon's********************* ** ** ********* emerging ***** *** ***** surveillance.

Comments (31)

I bet it works great in an empty boardroom with fixed lighting.

The investment bankers love it!

We have Apperence Search running on a 300 camera system and it does work with some limitations. Currently the cameras need to be able to detect individual people for the search to work. So right now if you have busy hallway (think school when classes are changing) the cameras can't detect individual people. Other times when hallways are not packed it works well. Also if you checked out the NVIDIA booth you will see Avigilon has a plan to address the density issue.

I would love to hear from the person that disagreed with my post.

The disagree button is a strange one. Ive had people disagree with me about my own experience. Apparently someone disagrees that you have a 300 camera system installed running appearance search?

Mike, why do you care about something so trivial?

I have no idea who voted to disagree with you but it doesn't matter. You must have better things to do?

Just be confident that by sharing your expertise the overwhelming majority of people will value and respect you.

I would love to hear from the person that disagreed with my post.

Why wouldn't you think it was the person that you were disagreeing with in your reply?

Thank you for the report and details on timing. It has been difficult getting a clear answer.

Is there a specific model of NVIDIA graphics card required? Assuming the correct NVIDA card is used does appearance search still increase CPU utilization?

Preferably GPU with Compute Capability factor of 5+

We are using the G Force Nvida 1060 6GB Ram and it works brilliantly. The site has 5.0MP H4A Bullet Cameras.

Note that the 4MB is the minimum. We tried with 2GB 1050 Card and it didn't work.

Hope that helps.

I would recommend using the card from Avigilon's price list (part # HD-NVR3-ANK1-1) as they might not support other cards if you have issues.

You need 768 CUDA cores Louis, try the GTX 1050ti.

You need 768 CUDA cores to run AS.

I hope IPVM will run own tests soon. AxxonSoft is also promoting forensic search by tracking, colors, faces and LPs and all works perfectly. I've demonstrated it personally to John on real cameras around the booth at ISC and hope to see tests at IPVM lab. This is very hot features and must be evaluated. Am I right ? :)

Update:

The Vehicle Search function is now available with ACC 6.2, available from Avigilon's Software Downloads Page

Regarding the 'Appearance Search' and other enhanced analytics. Info on the The "HD-NVR3-ANK1-1" does not show up when searched for on the Avig web site. Nice to know that it does not appear to be a proprietary video card.

What is a bit concerning is that our customers are reporting that the Avigilon rep is telling him (and other dealers) that if customers do not use their sourced server then they will lose out on some of the features built into the new Avigilon version 6, – vehicle search and analytics piece.

Folks use all kinds of servers, Dells, HP, Seneca (yes I work here ;-)), Lenovo, IBM that are not a Avigilon sourced servers. Is there any truth to what our customers are being told?

No Truth. Its all about the graphics card. We are running ACC6 with Nvidia Graphics card and AP works like a dream.

Once we install the Card, are there special software drivers we need to get from Avigilon to allow it to work (beyond the ones we can get from Nvidia)?

There's an analytics service which has to run on the server. You can find it on the support site. After that, there should be a check box in supported cameras to enable Appearance Search.

Ethan, the 'supported cameras' prompted a new question that I could not find the answer for since I have not installed this (yet).

Does this analytic require the use of Avigilon cameras only?

You need to use H4 Analytic cameras for appearance search to work.

Avigilon does have a configuration tool that makes some registry changes to enable appearance search but that information has not been shared yet unless you purchase the kit directly from them.

Avigilon support confirmed they are only able to support appearance search installations running an Avigilon rack mount HDNVR3/HDNVR2 and the M2000 needs to be purchased directly from them.

Min specs for the video card

Nvidia CUDA compute 5.0 or higher
Total Memory - 4GB GDDR5

(no mention about cuda cores but see specs for the M2000)

FYI, once the video card is installed, the server fan will crank up to 100% and it is very loud if not in a server room. This is being looked into.

The 'registry config' is a new item in this discussion.

This requirement is opposite of what Louis posted above.

If we can get him to expand on how he installed everything we can discover what we need to know. Louis?

Perhaps Avig would consider a 'license fee' if indeed the registry config tool is mandatory. I dont know what they are charging for the 'kit', but hardware wise, there is no magic there and folks will not be happy paying more than necessary.

If someone has an compatible video card, I would try enabling TCC mode on it. Just make sure it's not on your primary video card or you may not get back into Windows very easily!

http://docs.nvidia.com/gameworks/content/developertools/desktop/nsight/tesla_compute_cluster.htm

Avigilon has a new YouTube video demonstrating the UMD's particular search settings:

While it shows many configuration and display options, it is not a very compelling demo as the results they show do not seem particularly interesting or noteworthy. It's mainly surprising since demos generally err to the opposite showing unrealistic successes. We'll cover this more later but just flagging / adding this video in the existing post.

Not certain I understand the use case for this if it only functions on recorded video.

UMD is currently used for recorded video only to help find events faster. The idea is the system learns what is normal and flags what is not normal for operators to review.

Theoretically, I think UMD for live video has lots of potential. The problem is that it's hard to see UMD to being accurate or valuable enough for live video alerts to make it worth it, i.e., for every 1 UMD alert that people would care about 10x or more would be a nuisance.

I think it's important to emphasize the the difference between "successful imaging" and "successful analytics" usage.

Successful imaging (defining a scene view and knowing results for viewing and playback quality) is much different than analytics imaging (expecting accurate detection of desired events with minimal false positives).

In one scene (camera view), analytics of one flavor or another could be very accurate and successful. In another scene, analytics could be a complete failure (high false positives and/or lower probability of detection).

UMD (and other analytic detections) are going to have their sweet spots and not so sweet spots for usage.

The biggest headache in the real world is determining when a given analytic is effective for it's desired purpose. This is exceptionally difficult to quantify and will require real world deployments to substantiate effectiveness of the technology.

It's easy to know what a camera will "see." It's difficult to know what a camera will "accurately detect." In some cases it will be little nuisance, in some cases it will be medium nuisance, in s some cases it will be high nuisance.

The critical question is what nuisance rate is acceptable for the situation; 1x, 5x, 10x or more. It's rarely quantifiable and extremely situation dependent.

I have been testing Client machines that have multiple video cards in them for more performance.

What came to mind based on a customer query is if the AS function can also take advantage of multiple GPUs in a system to be able to support 'more'?

Login to read this IPVM report.
Why do I need to log in?
IPVM conducts unique testing and research funded by member's payments enabling us to offer the most independent, accurate and in-depth information.

Most Recent Industry Reports

History of Video Surveillance on Jul 19, 2019
The video surveillance market has changed significantly since 2000, going from VCRs to emerging into an AI cloud era.  The goal of this history...
Mobile Access Usage Statistics 2019 on Jul 18, 2019
The ability to use mobile phones as access credentials is one of the biggest trends in a market that historically has been slow in adopting new...
New GDPR Guidelines for Video Surveillance Examined on Jul 18, 2019
The highest-level EU data protection authority has issued a new series of provisional video surveillance guidelines. While GDPR has been in...
Anyvision Aims For 2022 Revenue of $1 Billion on Jul 17, 2019
Only 3 video surveillance manufacturers do a billion dollars or more in annual revenue - Hikvision, Dahua, and Axis. Now, Anyvision plans to join...
HD Analog vs IP Guide on Jul 16, 2019
For years, HD resolution and single cable signal/power were IP camera advantages, with analog cameras limited to much lower resolution and...
How To Troubleshoot Wiegand Reader Problems - Inverted Wiring on Jul 16, 2019
Wiegand is the dominant method of connecting access readers, but problems can arise for installers. In fact, one of the most difficult reader...
ZeroEyes Gun Detection Startup on Jul 16, 2019
A gun detection video analytics startup, ZeroEyes, is being led by a group of 6 former Navy SEALs, aiming to "save lives" by using AI to assist...
Motorola Acquires Watchguard, Adds to Vigilant And Avigilon on Jul 15, 2019
2 years ago, Motorola had no position nor relevancy to video surveillance. Now, they own major video surveillance, LPR and body camera providers...
Hikvision Global News Reports Directory on Jul 15, 2019
Hikvision has received the most global news reporting of any video surveillance company, ever, ranging from the WSJ, the Financial Times, Reuters,...

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.

About | FAQ | Contact