Why Verkada Is Moving Its AI From Cloud to Edge

By Sean Patton, Published Nov 06, 2020, 08:43am EST

Is the future of video surveillance AI the cloud or the edge?

For Verkada, the answer is clear - the edge, onboard its cameras.

IPVM Image

In a recent webinar, Verkada's engineering leaders explained why and, in this note, we summarize their position and analyze its approach.

Executive *******

IPVM Image

*******'* ***** ************* ** ******** ****** Yunchao ****, ***** ** *********** *** Cromwell, **** **** *** "actively ******* ** ******" AI *** ******** ****** analytics ** *** ****, which ********* **** ** Verkada ***** *******.

******* **** **** ********* have **** ***,*** ******* deployed *** *** "****** * ******* ****** a ***, ** ******** in *** ****** ** the **** ****** ** years ** ***** ** grow ** ******** ** people * ***" ***** would ********** ******** ******** Verkada's ***** ********** ****, which ** ******** ** the ****-**** ************ ***.

********, ******* ******* ********* stores ***** ** *** cameras, ** ** **** efficient ** ******* *** video ** *** ******, versus ****** ******* **** Axis ******** *** ******* Stratocast.

Webinar *********

*** ***** ***** ** Verkada's **** ************:

**********

*** **** ********** ** below:

********** ** **** *********

******* **** *** ******* reasons **** ** *** currently ***** ******** ****** analytics ** *** **** is ******* *** *** memory, *********** **** **** use "***** *** ********* of *** ** *** computer ****** **********" ** the *******, ***** ** limited ******** ** **** can ** **** ** an ********* ** ***** offering.

****: **% ******** ** Size, *** ****** ********

******* **** ***** **** is ** ******* ** models **** **** ******* 2MB ** ******, ******* of *****.

******* **** **** *** using ******* **** ****** network ************, ***** *** developed *** ***-***** ****** devices, ******* *************, *** ***** ******* filters *** ****** ********** times:

IPVM Image

** ******** ****** ******* pruning *** ************, **** reported **** *********, **** have **** **** ** reduce *** **** ** each ***** ** ~** megabytes, ******* * **% decrease:

IPVM Image

******* *** ************ *** both ******** ******** *** cost ********** ********/*********** *********, which ******* ***** ** took * ********* ******* for * ****** ** develop ** *** ***** using *** **** ********* architecture.

******* ******** ******* *********** advantage ** ***** ******* was ******** *** **** to ******* ** *****"**** ***** *** ****** to ***** *** ************"

*** ***** ** *******

******* *********** *** *** model *********** ******** ** using **** *** ****** OCR *************:

IPVM Image

**** ***** ****** ****/*********** versions **** *** ***** 2MB ***** ****, *** they **** ******* ** move **** **** **********.

Ambarella ********

******* **** **** *********'* n** ******** ****'* ****** out, **** *** ***** to ******** ** **** advancement ** *** ****, claiming *** - **** processing ***** ** *** neural ******** ** **-******:

IPVM Image

********* ********* **** **** that *** **** **** Verkada *** ************* ********* performance (*** - ***) over **** *********** *** that ***** *** ********* chips ***** ******* ********* in ***********:

*** *********** ** *** CV25 **** *** ****** AI ****** ****** *** S5L **** **** **** core *** *** **** is ******* *** *** 50x ***** ** ********* models *** ******* ******* video. *** **** *** twice *** *********** ** CV25. *** *** *** eight ***** *** *********** of ****.

******* **** *******

******* ******** **** ***** edge **** ** ********* small, *** *******, *** of ** ******** ********* (compared ** ***+ ***********), less **** ** ********* are ****-*******. **** **** while **** ***** ************ are ****** ** *** cloud, **** ****** * continued **** *** ******* to **** *** ** the *****:

... ,** *** **** and **** ** ************ going ** *** *****, so **** ********* ** the *****, **'** ******** starting ** ******* *** other ***, ***** **, we **** ** **** all **** ***** *** of *** ***** *** figure *** *** ** run ** ** *** edge.

******** - **** *************

** ***** **** ******* here *** *** *** edge ** *** ******* choice *** ***** **** video ********* *** ***** surveillance *******. *** **** benefit ** **** ** is **** ********* ** include ********* ********** ** a ***-**** **** ****** the ****** **** ** pay ******* ********** **** for ***** ******** ** the *****. ***********, **** also ********** *** **** to **** *** ***** to *** ***** *** processing, ***** ******* ********* cost *** ********* ***** of ******* ***********.

**** /****

Comments (27)

Undisclosed Manufacturer #1

11/06/20 01:58pm

**** ***********. *** "****** edge *******' ***** ********* to ******, *** * has **** ******* ** microsoft, ******, ***** (**********) in *** ****** ** edge *.*****. ** *** think **** ** **** the ***** ********* ******** everyone *** ******* ** reoccurring ***. *** *** "lower ***********, ** ********* on ****" *** ***** is ***** *********, *** all *** ********* ********** with "** ***** ******** grows, *** **** ** cloud **** ********" **** isnt *** ****. * major ****** ** *** amount ** ****, *** the **** ** **** systems **** ***'* ** process. (***** ***** *** processing ** $$$)

** * ****** **** compute ****** ***** *****

Undisclosed Manufacturer #2

11/06/20 03:35pm

* ****’* **** ******* I ******* ***** ** the ****** ****** ** a ***, *** **** the ****** ***** *** easy ********** ******** *** pushing ******** ** *** cloud *** ******* ***** more ******* ********** ** the ********, *** *** image ** **** ** the *****. **** ***** allow ******* ****** ** the *********** ** ****** the **********.

***** ******* ***** ************* to *** ****** ********* centrally ** ***** ** good ** **** *** algorithms ************ *** *******.

****

Undisclosed End User #3

11/06/20 04:15pm

* ** ******* *** for *** **** ********** of ******* ******* ** Intel ********, **** **** be * **** ******* for ***/************ *** *** what *** ****** *** do **** **** *** away. ***** * ****** with ** **** *** its *** *********, ******* traditional ********* *********.

John Honovich

IPVM | In reply to Undisclosed End User #3 | 11/06/20 04:18pm

#*, * *** ***** ago, * ******* **** Intel / ******** / Myriad / ******** ***** were ***** ** ** big ** ***** ************ but * **** *** seen **** ** ********. From ********* **** ********, Bosch, *** *******, ** name * *** ****** players, ********* *** ******* won. ** ************* **** manufacturer ******** ** **** this ** ********* ******* Ambarella ******* **** (**** video ********** *** **) while ***** ****** ***** are **** ***** **. If * ** ***** about ****, ****** ** invited ** ******* ** and ******* ***, ******!

Undisclosed End User #3

In reply to John Honovich | 11/06/20 06:51pm

** * ******** *** Intel ****** ****-*** **** the ********? * **** they **** ** **** blown ** *** ***** go ****** *** * still ***** **** **** the **** **** ** the *** ***** **** it **** ******.

John Honovich

IPVM | In reply to Undisclosed End User #3 | 11/06/20 09:37pm

* ***** ***** **** once *** **** **** in *** *** ***** then ** **** ******.

*****, ** **** ***** one *** **** **** case *** ******. **** are * **** ******* and * ***** ****** in **** *** **** are *** ****** *** analytics ********* ** *** edge.

*** ******** *** ***** and ****** ** ** they ****** **** ***** winning *** **** ****** analytics ******. ** *** just ** *** ***** for **** ******** ** the ******* ************* **** have ** ****-******* **** and ******* ***** ************.

Michael Miller

In reply to Undisclosed End User #3 | 11/06/20 11:22am

***** *** ******* ******* with ******* ** *** LTE ** ****.

** ************* ** *********** heat *** ***** ** also ** ***** **** the ******** / ****** / ******** *****.

Avatar

Brian Karas

11/06/20 04:28pm

*'* *** **** **** the **** ********** ******** all *** *******.

***, ********* *** ****** to **** ***** ****. The **** *** ** in * *** ** different *********, ********* ** a ****** ** *******, but *** *** ******* is **** ** ***** to ** **** **-***** object *** ***** ********, GPUs *** ***** ** be "*** *****" *** the **** ********** ** analytics.

** **** ****** ********, one ***** **** ** heat/power (*** **** **** the ******* **** ** an ******** *** **** fits ** * ******).

* ****** ** **** see **** **** ** Moore's-law **** ** ******** with ****, *** ******** not ** *** ***** that **** ****** *** size/power ********** ** * modern ***, ****** **** will *** ********** ****** as * **** ******.

**, ** ******** ** that ** *** ***** to *** ********* ******** to *** ***** *** the "*******" ***** ** analytics - ******* ***** (though ***********), *** **** importantly ******** ****** ** lots ** *****, ********* easy/secure ****** ******, ***. But *** ********* **** be **** **-*******. **** likely *** ***** ** the ******, ** ***** on * ******** ****** or ********* ***, ** maybe ** * ******** of **** ****. *** answer **** ******** ****** on *** **** *** the ******** (************) **** choose.

** ******* **** ******* about *********, ****'** ***** to **** ** ******* some **** ** **-**** unit ** ****** *** heavier ******* **** **** you *** ** ** a ******, ***** ** going ** **** **** closer *** ****** ** traditional *************.

Avatar

Morten Tor Nielsen

prescienta.com
In reply to Brian Karas | 11/06/20 06:15pm

**, ** ******** ** that ** *** ***** to *** ********* ******** to *** ***** *** the "*******" ***** ** analytics - ******* ***** (though ***********), *** **** importantly ******** ****** ** lots ** *****, ********* easy/secure ****** ******, ***. But *** ********* **** be **** **-*******. **** likely *** ***** ** the ******, ** ***** on * ******** ****** or ********* ***, ** maybe ** * ******** of **** ****. *** answer **** ******** ****** on *** **** *** the ******** (************) **** choose.

*** **** **** * years **** *** ******, this ***** **** * description ** * ******* high *** ************:)

*** **** ** **** today *** ****** *** choices *** ********* *** analytics ** **** ******* a **** ****** ** unnecessary ********* *** ** rendering *** ***-******* ***. that ** **** ***'* need - *** ***'** paying *** ** *******. As *** ** ** is, ** ** **** we **** *********.

** ******* ********* ** the **** *** ***'* need ** ****** ***** prior ** ********. *** also *** ** ** the ********* ** *** lowest ******** ******* *** at ************ *******. ** what ***'** **** ** a ********* **** ******* of ******* * *** at * ****** **** (tensorcore).

** ****** *** ********** cost ** *** ******** being ***** ** *** camera, ** ***** **** the ****** ** **** more ********. ******* ** building * ********* ********* super ** ** ****** 100 *******, **** ****** brings *** *** ****** to *** ******. ** the ***** ** ****, you're ********* *****. **** 100 ****** *******, ** is **** ******** ** one ** *** **** at *** **** ****.

***** **** ****** ** a **** *** ***-**** and *****-***-**** ********. ********* may ****** ******, ******, colors *** ** **, but ** ***'* **** everything **** *** ***** want ** **** ***. In ***** ***** ** don't ****** **** **** choice - *** * beefy "********* *********". **** analysis *** ******** ** accelerated ** ********* ***** swaths ** *** ******** by ******* ** *** metadata ********* ** *** cameras **********.

* ***'* *** ** much ***** ** ***** recording; ********* **** ******* a ********** ***** ***** is *** **** *** of * **** *** medium *************, ******** * lot ******* ***. *** micro-installs, ***** ******* ***** sense (******** **** ** manageable).

**** ******* * ***** edge ***** ********, **-**** recording *** ***** ***** infrastructure ********** ** *** optimal *** ** ***** things *****'* **** **** this ** *** ****** will *** ***.

* ***** *** ** need ** ** ***, is **** ** *****, and *****, **** ***** ONVIF **** **** * 200 **** ******** ********** a ******** *** ******* for ***** ******.

Avatar

Brian Karas

In reply to Morten Tor Nielsen | 11/06/20 07:22pm

*** **** **** * years **** *** ******

* ****. **** ** what * ** **********. It ** ****** * upgrade ***** ** ********.

** ******* ********* ** the **** *** ***'* need ** ****** ***** prior ** ********.

* **** **** ******* familiarity **** ********* ** the ****. **'* ***** for *****-***** ***** (**** is * ******, **** is **** **** ** vehicles), *** **'** ***** a **** *** *** (>1 ******* ***** ** least) ****** *** ******* edge ****** *** *** "This ** * **** with * *** *****, blue *****, ****** ** an ******* *******" *** do **** *** **** for * ** * objects, *** *** **+.

*** *** **** * years, * ** *** see ********* ******* ******* to ***** ** ** what *** ******** ** doing ***** **** ***/******-***** systems. * ** ****** to *** ******* *** smarter *******, *** * expect ** *** *** GPU/server ***** *** ******* and *******.

Avatar

Morten Tor Nielsen

prescienta.com
In reply to Brian Karas | 11/06/20 08:20pm

*** *** **** * years, * ** *** see ********* ******* ******* to ***** ** ** what *** ******** ** doing ***** **** ***/******-***** systems.

* ***** **** **** - *** ******** *********** of *** ****** **** always ** ******** ** what *** *** ** with * ********* **. No ********.

*** ** ** ***'* see ******** **** ******** I ***'* ***** ** will ** ******* * monolithic ******** **** * GPU ** ******** (*********** cost). ** **** ** because ***** ** ****** or *********** ** **** ********. The ******** **** ***** the *** *** **** show ********* ** *** of ** ***-****** **/*** pair, *** **** ***** a *** ******** **-*** with * ********** ** dumb *******.

** * ***** ** some ***** ********** *** I ***** **** ***. I'm **** ****** *** as ** *** ******.

John Honovich

IPVM | In reply to Brian Karas | 11/06/20 09:34pm

*** *** **** * years, * ** *** see ********* ******* ******* to ***** ** ** what *** ******** ** doing ***** **** ***/******-***** systems.

*** ****, ** **** you **** ***** ** the ******** '****' / most ******** *********.

** *** ***** ****, it ** ******* ******** apparent **** *** **** common ********* *** ************ being *** ********** ** the **** (**** ** people / ******* / face *********, ***.).

Avatar

Brian Karas

In reply to John Honovich | 11/06/20 10:57pm

**** *** **** ****** analytics *** ************ ***** run ********** ** *** edge

**** ** (******) ***** done ** *** **** today **** **** *********** is ****** ********* **: "this ** * ******". What ** *** ***** done, ** **** *****, is ******** *** ********* down *** ******* (****, red *****, **** *****, has ***, *** ***). High ******* ******** ** the **** ** *** or ****, *** ** a ******* *** **** proven ** ** ****** (though *****-****** ******** ** the **** ** * fully ****-***** ****** ****** to *** *****, ********* that ** *********** ****** done *********).

** *** ******* ** that ********* **** ****** be **** *** ***** alerting (****** ******** ********, vehicle *********), *** *** storage ********** (**** ****** video **** ***** ******* present ******* ** ***** motion), **** **** ********* make * *** ** sense.

*******, * ***** *** slow *** ********** ****** of ********* ** *** edge ** ******* **** awareness *** ****** ** a *** **** **** push ** **** ******* deeper ***** ** *********. By **** * ****, using ********* *** ******, doing **-************** ****** (**** other *********** ** **** object) *** **** **-************** alarms (**** ** *** next **** *** ****** who ***** **** **** comes **** **** ** a ******). ************ ** the ****** *** *****, just ********* *** ****** or ******* **** ****** quickly. ***** **** ** search *** * ****** SUV, ** * ****** that **** ** ******* path ******* ** **** is *********** **** ********, and **** ****** ******* a ***** ** *********** that ** *** ********* at *** ****, ** will ******* ********* ******** FOVs, ****** ** **** logical ** ** **** in **** *********** ******.

* *** *** ***** increasing, *** ** ********** in *** **** * years * ** *** see *** ****** ******* getting *** ****** **** we *** ****** *********** surveillance **** ** *** mainstream. **, ** ****** points ***, **'** ***** using **** **** **** with **** ********** *******. To **, **** *** points ** *** ****** todays ***** ** *** compute ********* ** ** SoC **** ** ******** for ***** **** ** a ************ ****** *** driving **** * ***** call **** ***** **** analytics. *******, ** *** squeeze **** *** ** those ********** *********** **** by ******* **** ** the ***** ********* ***** to *** ****. *** the ****** **** *** the ****** *** ********, and *** *** *********** GPU ** *** ****** analysis, ******-**** ******** (**: unusual ******* *********, ***** analysis) *** *****-****** ************ (show ** *** *** FOVs ***** **** ****** appeared).

******, ** ***** ** additional *************, ** **** of ** ***** ************ come ****, **** **** we **** "*********" ** likely ** ***** **** basic ****** ********* ** deeper-level ********, ***** ** why * ** *** think "*********" **** ** predominantly **** *****, ****** on * ***** ****** count ***** ** ***** have **** *********** ******* than *********** ****, ***** would *********** **** ******* the **** ****** ***** for ********* ** *** in * ***** ** we ***'* ****** *** definition ** ********* ** advance ****. ** * counter ******** ** ** own *****, ** ** get * ***** ****** Law ***** ** *** advancement, ** ***** ****** more **** ********* ** add ***/******-***** ********* ** existing ******* (******* ** mind * ***** ** barely * ******* *****, lots ** *******-********* ****) making *** ******-***** ********* the **** ****** ** 5 *****, *** ***** not ** ** *****...

John Honovich

IPVM | In reply to Brian Karas | 11/07/20 03:22pm

******, ** ***** ** additional *************, ** **** of ** ***** ************ come ****, **** **** we **** "*********" ** likely ** ***** **** basic ****** ********* ** deeper-level ********, ***** ** why * ** *** think "*********" **** ** predominantly **** *****

**** ***** **. ** gut **** ** **** this **** ****** ****** in *** ********** *****, like *** ******** *** been **** ** *** past ******. * ** not ****** **'* *** valuable, * ** ***** it's ******** *** ***** the **** ********* ** enterprise *****.

****'* *** * ***** the **** ****** ******** of ********* ***** ******* will ** *** **** with *** ***** ** server *** ********** / government / ******** ************** users *** **** ** do ************* ******** ** analysis.

Undisclosed Manufacturer #2

11/06/20 08:13pm

** ******* * ***** with ******. *** ******* for *** ****** ** process ***** *** **** it **** ****** ******** will ** *** ***** step *** ********.

Kris Tibbitts

11/08/20 04:10am

****, ** **** **** an *** **** ***** not ** ** *****. Pack ****** ** ********** in ** *** ** problem. ***** ****, **** easier ** ******* * single *** **** * bunch ** *******. *** then **** ***** *** make ** **** **** reason ** ****** * times *** ***** **** the ******* *** **** worth, **** * ************.

Undisclosed Manufacturer #2

In reply to Kris Tibbitts | 11/08/20 05:14pm

**** ***** ***** *** part ** *** *****, and *’* *** * Verkada ***. ***’* **** this *******.

*** *** *** ** manager *** * ****. You **** ** ****** the **** ******* *** that **** ** * change **** **** ** IPv6 ** **** ** new ***’* *** ** schemes.

***** *** *** *********, 6,000 ******* *** ***** a ******** *******. *** of *** ****** **** be * *********** *** facial *********** *** *** on ******** ***** *** only ******* ***** *** browser *** ****** **** have ****** ** **** and *** **** ** log **** **** **** accessed ** *** ** able ** *** * report *** ***********.

**** *******, ** ** works ** ****, **** would ** **** **** seamless.

*** ***** * ***** of ***’* ** **** multiple *** ********* **** in **** *********** *** what ***** ** **** in ********, ********* *** time?

*** ** ******.....*’* *** a *** ** *******’* current *** ** ***** offering.

John Honovich

IPVM | In reply to Undisclosed Manufacturer #2 | 11/08/20 05:54pm

** *** *** *** IT ******* *** *** want ** *** ** other ******* ** ******* an ******** ****** ******* or *** ** *** in **** ***** *********, etc *** **** *** are *******.

* ***** **** ** their ***** ***** ** that **** **** ** easy ** *** ** and ******** ** *** out ** *** ***** things.

Undisclosed Manufacturer #2

In reply to John Honovich | 11/09/20 10:29am

****** **** ********?

Michael Miller

In reply to Undisclosed Manufacturer #2 | 11/08/20 05:56pm

**** ** ************ *** working ** ******* ******-***** solutions **** **** ** able ** ** **** without ****.

Undisclosed Integrator #4

11/08/20 01:44pm

******* *** **** ** edge ****** ****** *** years.

John Honovich

IPVM | In reply to Undisclosed Integrator #4 | 11/08/20 05:55pm

***, *** ******* ** selling **** **** **** turnkey **** *******, **** in ***** ** ********** / ***** *** **** sales.

Michael Miller

11/09/20 03:22am

***** **** ******* ** adding *** *****.*****

IPVM Image

Avatar

Kenneth Kocher

Force Security Solutions
In reply to Michael Miller | 11/09/20 11:06am

** *** **** *** a *** ** *** customers *** **** ** very ****. **** ** great ******** *** ********. This ** * *** party *** ** ******* actually ******** **-***** ***** is * ** ***** Eye ******* **** ************ for *******. ** *** currently ******* **-***** ** this **** *** **** like **. *** **** is **** ** *** use **** *** *** and *** **** *** Eagle *** ****** *** the ******* ** **** instead ** *** ******.

Avatar

Christoph Derndorfer-Medosch

Netavis Software GmbH
In reply to Michael Miller | 11/09/20 01:00pm

** **** **** ***** is * ******** ****************** *******, *** **'* not***********.

John Honovich

IPVM | 11/09/20 06:49pm

*******:******* *** *** ********* LPR.

Michael Miller

In reply to John Honovich | 11/09/20 06:53pm

**** *** ***** ******** is ****** ******* **** use **** *******. **** other ******* *** **** cameras.

Read this IPVM report for free.

This article is part of IPVM's 6,731 reports, 908 tests and is only available to members. To get a one-time preview of our work, enter your work email to access the full article.

Already a member? Login here | Join now
Loading Related Reports