Video Analytics Hardware 101

By IPVM Team, Published Feb 16, 2021, 10:43am EST

Video analytics can be supported on different hardware types and programming frameworks, impacting cost, efficiency, and accuracy.

IPVM Image

The primary factors reviewed in this report are:

  • CPU including Intel
  • GPU including NVIDIA, CUDA and Tensor Cores
  • VPU including Myriad / Movidius
  • SoCs including Ambarella and Huawei
  • Tradeoffs between using CPU, GPU, VPU and SoCs

This is part of our new Video Analytics Course starting in March.

4 ********** *******

*** ** *** ******* limitations ** ********** ***** analytics ******** ** ********* processing *********. ***** ********* are **** ******** ********* in ***, ***, ***, or ***.

IPVM Image

Executive *******

*** ********* ********* *** tradeoffs ** *** * options:

  • **** ** ******* ** recorders *** **** ********** used *** ***** ********* because ** ********** *********** but ***** *** ******** OpenVINO ** **** ** its ******** **** ** optimize ***** ******** **********.
  • ****, **** ******** ******, are ******** **** ** servers ********** ***** ********* for ****** ****** ******* of ******* ** ******* analytics.
  • ****, **** ******* ***** Myriad, *** **** ******** for ***** ********* ** edge ******* **** ** IP ******* *** **** gained ******* ******** ** date ****** ****.
  • ****, **** ******* ********* and ****** *********, **** become *** **** ****** choice *** ***** ********* in ** *******, ***** the ******** ***** *** simplicity ** ******* ***** encoding *** *********. *******, Huawei's ********* ********** *** been **** ****, ** least *** ***, *** to ** *********.

Processing ** ***

**** *** *** **** widely **** ********* ********, generally, *** **** ********* offer ******** ***** ******** performance **** ****.

IPVM Image

***** **** *** ** used ***-***-***** ** *** video *********, **** *** only ******* *** ***** rates *** ******* ******* of *******. ****** **** depends ** *** **** of *********, *** ********** of *** **********, *** the *** ****, ** is ********* *** ****-********* to *** ***** ************ analytics ** ****.

*******,*****'* ************ ** ******** *** performance ** ***** ********* on ******** ***** **** (e.g. **** **/** *** Xeon **********). ***** ***** surveillance ********* ****** ******* Intel ****, **** ******* cost *** **** ******** to ********** ******* *** other *******'* ****. *******, while **** ******** ********* are ***** ******** **** CPUs, ** ** *** widely *******.

Processing ** ***

**** *** ******** ****** for ********* ** ******* because **** *** ******** to ******* ******** ***** processing **** ******** ** thousands ** ***** *** GPU ****. **** ******* a ******** *** ******** providers ** ****** ****** networks *** ***** ***** and ******* *********.

IPVM Image

************, ***** **** *** analytics ****** *** ****** system ***** ******* ******** cards *** ** ***** to * ****** ****** to ******* ******** ** cameras, ********* ** *** complexity ** *** ********.

*******, **** **** **** for ********** ***** ************ cost ********* ** ******* per ****, ****** **** to ***** ********* *** lower **** **** ** only ******-***** **** **** used.

NVIDIA **** *** ****** *****

****** ** *** **** commonly **** *** ******** for ***** *********. ***** NVIDIA ****** ******** *** families, **** ***** ******** providers ******* ******* *** and ****** ****. ******** cards ** ***** ******** typically ***** ********* ** CUDA ***** *** *** generation *** ***** ***** hundreds ** ****** *****, at ***** ******* **** $400 - $*,*** *** card.

****** ****** *** ***** with ************* ****** **** and ****** **** ****** (***** *,***** ****) **** ******* ** Quadro, *** **** ************ applications ***** *** *** the ********** ********** ***** and *** **** **** ($3,000 - $**,*** *** card) ** * *********** drawback.

***** ******** ****, ******'******* ************** ********** *** **** AI ********* **********. ***** they *** *** ******** used ** ***** ************ analytic *********, ***** *** cost ($** - $***) for ******** ** **** cores ** *********** ** AI ******** *** ********** per ****** ***** *** low ****** *****, ******** systems.

****** **** *** ****** used ****** ** ** cameras ******* ** *********** on ******** ****, ****, and ***** ***********. ****** has ****** *** ** compete *** **** ****** of ** *******. ** contrast, **** *** **** have ******** **** *******.

Processing ** ***

***** **** *** *****-**** than **** *** *** be **** *** ********** video *********, **** ** not ******* ***** ********, putting **** ** * disadvantage ** ****. *******, VPUs **** ******** *** AI **** (*****/******) ********, offering ******** ********** *** higher *********** **** ****.

IPVM Image

***** ****** (******** **** Movidius) ** *** **** common **** ******** *** video ************, ******* ** server-based ***** *** **** as ******* ********** *** cameras. ******* ***** **** support *** **** ********* as ***** ****, ******** providers *** *** ******* for *** ************ ******* with ******* *********** ******** to ***** ****.

*******, **** *** *** getting *********** ******** *** analytics **********. ****** ************* have **** ********* **** that, ****** ****, ******* both ***** ******** ***, more ********, ***** ********* processing. ** ****, ** is ********* **** *********, both ** ***** ** forgoing *********** ***** *** adding ** *** ******* and *** ******** ***** encoding **********.

Processing ** ****** *** (****** ** * ****)

*** ** *******, **** are ******** ****** **** CPUs ** **** ***** they ******* ***** ******** into * ****** ****** that ********** *** ******* the **** ** **** encoding *** ********* *** video.

IPVM Image

*** ********** ***** ** steadily **********, *** **** modern ** ****** **** are ***** **** *********** image ********** ******* ***** make ****** ******** ********* possible, *** *********** ****** widely (*** ******** ****** ********* ********).

Huawei *********

*********** ************ ************ **** *** most ****** *** **** in ** *******, ********* performance ****** ************* ******* models, *** ********* *** not ******** **** *** most ******** **** *****-**** chips.

*********'* ****** **** *** most ******** ****, ******** a ***** ** ***-****, limited ********* ****** (******* *********), ** ******** **, multiple ****** ******* ********** versions (****** *********).

*******,********* ** *** ********* manufacturing ****** * ****** ** US ***** ********* ******* its ****** ******. ** such, ***** **** ************* have ********* **** *********, unless ********* ******* ** the **** **** ** two (***** ** *****), Hisilicon **** *** ** a ********* ****** *** IP ****** ****.

Ambarella

**-************************ * ****** **** SoC ************ *** ** cameras ******** ***** *********. Ambarella *** * **** product ******** *** ***** surveillance; ** *** *. The ** ****** ** for "******** ******" ** video *********-******* ********. *** S ****** ** ******** for ******** *** ********** IP *******, ******** ** low-cost ********* ***********, *** limited *********.

Manufacturer ***** - ****** *** ****

** ******** ** *** broadly **** **** **** Ambarella *** ******, *** video ************ ************* ******* their *** ***** - Axis (******) *** ****** (*******). ***** ************* *** positioning ***** ***** ** deliver ***** *********. *******, even *** ***** *************, they *** * ******* of ***** *** ********* chips *** **** *** parties (*.*., *********).

Challenge ** ***** ****-**** ** ****

**** ********* ****** ** using **** ** *** video ********* ** ** cameras *** * ****** for *************. ***** ***-** SoCs **** **** **** $10, **-******* **** ***** increase *** ***** ** 2 ** **, ********* on *** ** ************ and ****** ********** ** the ************.

** ******** ** $** for *** **** ** a ****** ***** *** seem *********** ***, **** marked ** ** *** manufacturer, *** ***********, *** the ******, *** ****** in * ***** ***** increase ** ~$*** *** camera.

Processing ** ******** ***

************ ********* **** *** SoCs ** ******** **** and ******** *********** ******** to ****/****/****. *******, ******* recorders *** ********* ***** from *-** ******* **************, the ********** ***** ** shared, ***** ********* ******* accuracy ** ************ ********* to ******** ********** **** fit *** ********** *********** of *** ********.

IPVM Image

************, ******* ********* *** generally **** ** *****-********* systems, ************* ** *** want ** ****-******* *** hardware ** **** *** recorder's ***** ***, ***** generally ***** ********-***** ******** is ***** **** **** other *******.

*******

** ****** **** ********* will **** ** *** SoC ******* **** ** the ******* ***** ************ manufacturers **** ******* *** including ********* ****** *** camera ****** ********* *************** and ********* ******* *******.

************, *** ************* *** increasing ********** ***** *** neural ******* *******, ****** by ******** ****** ************ outside ** ***** ************ (e.g. **********, *********, **********, medicine), ***** **** ******* IP ****** **** ** well.

Comments (24)

*’* ******* ******* ** Avigilon *** ********** *** party ***** ******* **** the ***** **** **** in ****, **** **** themselves *** *****. ** this ****** ******** *** VMS ********* **** **** sell ******** ** ***** integrators **** ****** ******** from *** *******? * thought **** ************ *** just ****?

* ***** *** ********** of "***********" ***** *******. I'm ****** **** *** Avigilon, "***********" ***** *** can't *** **** ** an ***** *** ***** tech ******* **** ** sales ********* *** ******* more. ** *****'* **** it ***'* ****, ** only ***** **** *** are ** **** *** to **** ** ****. I ***'* ***** *** VMS ************ **** "*******" a ***** ***** ***** of ******** ** *** didn't *** ** **** them. ************* **** ********* "support" *** ******** **** sell *** *** **** more.

** ** ********** * non ******** ******* *** run ***** ********* ******** for ********** ******, *** they ** ** *** of ***** *** ** make ** **** ********* if *** *****'* ********* their ******** ******* **** have ******** ****** ******** that *** *** **** get **** *** *** the ******* **** ****. Otherwise *** *** ******* the ******* ******** ** make *** *** **** and **** *** *** the ******** ** **** as ***** ********* *******.

* ***** *** *** this ** ****** ******** and ** ** * little ***** ** ** opinion, *** ***** *** ways ****** **, **** are **** ****** ** less ********** ** ** more ********* ** ******** their ********.

*** ***** ******* **** bring ** ** ***** it ***'* ***** ****** to **** **** ******* in * ****/****** **** factor ** **** **** have ******* ****** *** engineered *** ******** **** Dell ** ****** *** won't ** ******* *** residual ******** ******, *** I ***** *** *** how **** ****** **** actually *******.

* ***** ****** *** party **, **'* ** own ****.

* *** *** ****** ACC ******* ** *** party ******* ********.

***** * *** ** emphasis ** **** ***** processing **** **** *********** and ********** *****. ** the ** ************ ****** larger **** **** ***’* these **** ******* *** out ** ********** ***** rendering **** * **** investment ** ***** ************* to *** **** ** again ******* ********? ** it *** **** ****** to ******* ** *** cloud?

***** ******!

* ** ***** **** edge ********** ** **** beneficial *** **** ******* you ******* *** * think **** ************* **** went **** ***** ******* it ***** ***** **** devices **** ********, *** I ***** *** ****** will **** **** *** benefits ** **** ********** don't ******** *** **** term ******** ** **** and ****** **** ********* to *** **********.

**** ***** ****, * don't ***** ***** ** the ****** ****** *** because ***** *** ***** too **** ************** *********** like ***** **** ** stream ***** **** ****** a *** ***** **** creates ****** **** *********** of ******** ******* *** packet ****.

* ***** ***** *** most ******** ******** ** a "******" ***** ***** the ********* ********** ******* on ****, *** *** easily ** ******** ******* replacing *** ** **** cameras/devices. **** *** ***** can ** ******** *** it's ******* ********** **** multi **** ********** ** data *********/*******.

**** ** *** * see ** **** ** experience

*** *** ***** **** depend ** **** ************ and ******** *****?

*** **** ********** ** the *** *******...** *** know *** ****** ****, space ********, ***** *********** and ***** ************ ** power **** **** ** equipment? *** ***** ****** its ************...** *** ******.... i **** **** ****** that ****** **** **** brought *** ********** *** ****

** ***** ** ****** the ****** ***** *** currently ** *** ****** are ***** *********, ** just ***** **** ** the ****** *********

"** *** ** ************ become ****** **** **** won’t ***** **** ******* run *** ** ********** power ********* **** * lost ********** ** ***** manufacturers ** *** **** to ***** ******* ********? Is ** *** **** easier ** ******* ** the *****?"**** ** ******* *********... every * ***** *** manufacturer **** *** ******** patch **** ******* *** to ***** *** ** life *** ********* ******* a ************* ** *** network, *** **** ** replace *** *******.

** ** *** **** easier ** ******* ** the *****?

***, '******' ******** ********* (since ** ********** *** on-board ********** ****) *** it's *** **** ********* generally. ** ******* **** here -***** ********* ************ ********* ******* ** ****** Its ** **** ***** to ****

******* **** ****** *** same *** **** ****** phones ****...

*** ***** **** * mobile ***** ******, **** we *** **** *** basic ****, *** ** have * ***** ********** computer ** *** ***** (iPhone, ******* ***)

** **** ******* - first **** *** ***** camera **** **** ***** and ****, **** ***** audio **** */* ***, then ******* ******* (**), next ** **** ** a **** ******* ******* device **** *** ******* to ******* ******** **** we **** ** **** the ******** ***********.

* ** **** **** looking ******* ** ****** what’s ****** ** **** space...

****’* ****** ** **** space

******* ******, *******, *** yet **** *******.

** *** ** ***** systems ** ****** ******** of ***** **’* ***** deployment? ** *** **** just ********* * ****** NN ********* ** ***-****?

** ***

** *** *** ******, if **** *** **-********** the ****** ******* ** the ******, ** (****** I ***** ***** *** word '***' ***** ***** there ** ******* ** the ***** *** ** that ****, **'* **** rare). ******** ********* ***** far **** ********** ***** than ******* *********. *** alternative ** ** ******* the ********* ********* *** then ***** * ******** update **** * ***** trained *****.

******** ********* ***** *** more ********** ***** **** running *********.

***, ****’* ****. ****, the **** *** ***** have ******* ** ********** expansion ** *** *****/***********, that ***** **** ***** blush.

************, ******** ***** **** the ***-**** ****** *** available *** * *** hundred *******. ******** * took ** ******** ********** to ******* ******** (****-*** video) *** ** ****** hardware *** **** **** impressed. ******* **** ************ are *** *** **** from ***** **** ********* for ******** *****.

*** *********** ** ** retrain *** ********* ********* and **** ***** * firmware ****** **** * newly ******* *****.

******. ***** ******* * think **** *** ********** exists **** ***** ******** learn ** *** ***, specific ** *** ***** inputs, **** **** * human *****, ***** ***** increase ******** ** ***** systems *******. ***** **’* apparently *** *** *** case, ** *** ** worth **********.

***** ***** ** ******* in *** ***** *** does...

** ***...

**** ** "********"? **** that * ** ***** of. ** ** ***** far *** *************** *********, and *********** *** **** of ***** ******* (************ large ******, **********/******* ****, etc.).

**** ** *** ********* that ********* * ******** function *** ******** ****** to ** ****-*********** **** actual ********.

***** ** * **** recent ***** ******* ***** some **** ******** ****** in ******** *******, ***** false-positives *** ********** (******* via **** *****) *** used ** ****** * mini ******** *** ** help ****** ****** ***** positives. ** * ***** this ***** ** ******* of ** ****-********** ********, as *** ******** ******* basically ******* ** ****** class **** ** ******** objects ** ** *******. Like ***** ****** *******, if ***** ******** ******** examples **** * *** of **********, ** *** be *********, ** **** are *** ****** *********, it ***** *** **** much.

******, *****, *** **** great *********** ** ******.

******** *** ******** *** real *****, *** **** bears **** ****. ***** is * **** **** vignettes ********** ** ***** alarms *** **** **** similarity ** ********** ******. So, ** ****** ** example **** *** ******** domain, ** *** ***** a ************* ****** ******* to **** ***** ** a *****'* ***** *******, then *** ******* ******** training ** ******** ***** alarms, *** **** ** absolutely ******* **** *** of ***** ****** ********** for ******** ******** ****** do *** **** ***** hidden ****** ****. **** is ************ ***** *** edge *****, ***** *** example *** ****** **** like ******* *** * small ***** **** ** the ******** *** *** recognized ** * **** wake.

****** *****, ****, *** this ***** *********** *******!

******* ************* ****** **** space *** ****** **** exciting. *'** **** ********* appearance ********* ** ******** nighttime ***** ****** ****** by *********** ********** *** stray ********. *** *** provides ********* ******** ****, with ********, *** ** pieced ******** ** **** than ** ***** ** codes.

****** * ********** ****** of ********, ***** *** somewhat ********* ****** ******** frames. *** *** ***** rate ** ******** ** consumer * *** ** not **** ********* **** dual ********* ***** ** storage *** ** *** analysis ******.

*******, ****** **** *****, I'm * ****-*******. * know **** ****** ***** video ************, *** ***'* work ****** *** ******** or ************ ********. * only ******* ****** ** help ** **-****-*** **** his ****** ********. **** profound **** ** ********* is * ******* **********. I **** * ******** (classification ** ******* ****** surveillance ***** ** ******* 2 ****** * ******) that *'* ********* ******* as * ****-********* **** in * **** ****. I **** ** ******** parse *** **** *** objects ** ********, **** zoom ** **** **** within ****** *****. ** industry ******* ******* **** metadata ****** *** ***** itself, ******* ** *** metadata *** ** ****** in *** ***** ******? Is ** ***** **** proprietary ********** ****** ****? Has ***** ***** ********** this ****? *** **** open ********* ********* ** emerge? **** ** ** area **** ***** ** interesting ** **** **** about.

****** *** * ***** web **** **** ** accessible ** ** ********!

*********! * ***'* **** to **** **** *****, but * *** **** you **** ******* ********* my ******** ** ***** Analytics ************ ***, *** ONVIF ******* *:

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

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

*** ***** *** ** the *** **** ** in ** ******* **** interesting "******* *"

***** ********** ******* * Release ********* *** ******** and ********* *** ***** Applications - *****

*** *****

******* * - ******* Candidate - *****

******* ** **** *** the *********** ** ********* some ** *** ***** "smart" ****** ** **** some **** **** ***** solutions.

*** ******** ** **** take **** **** ****** it **** ***'* *** cameras *** ** *** hope *** ***** ********, to *** *** **** "independent" ********* ********** 😁

** ****?

*** ** ***** **** for ***** ************ ********* currently? *** ***** ** not ******* ** ***** be **** *** ******* it ** ***** ****. FPGAs ** *** **** to ** * ****** choice *** ***** ************ analytics.

****'* **** **** ********* benefits **** ***/***, ************ in **** ************ (******* power ***********), *** **** have * ***** ******** curve ** *** *************** side **** *****. *********** how **** ********* ********* start **** **** **** of ****-****** ********* **. a ******-** ***********, * don't ***** **'** *** many ****-***** ***** ********* in *** ******** ******** any **** **** (** maybe ****).

***** ***** ** ***** something?

IPVM Image

** **** ***....

***** *** * *** companies ********* ***** *** machine ******, **** ** Xilinx (***********). *'* **** not **** **'* ***** anywhere *** ******** ** the ***** ****. ***** is ********* ****** ** fight *** ****** ******.

Read this IPVM report for free.

This article is part of IPVM's 6,805 reports, 913 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