Video Analytics Algorithms / Efficiency

By IPVM Team, Published Apr 09, 2021, 10:58am EDT

One of the biggest barriers to video analytics is having sufficient computational resources. An alternative to increasing hardware performance for video analytics is improving algorithm efficiency.

IPVM Image

** **** **-**** ****** we *****:

  • *** ********* ********** *******
  • ***** ************ ****-****, ********** Intensive
  • ********* ********** *******
  • **** ******** ************ **********
  • *** **** **** **** CNNs
  • **** ********/***********
  • ****** / **** ******** Even **** *********, **** Accurate
  • ******** **********
  • ******* ******* *** ************
  • ****** **. ***** ********
  • ********* **** ******** *** Machine ********
  • ******** **********
  • ************ ********* ******
  • ********* ******** ******
  • ****** **** ******** ***** To *******
  • *** ***** ** ********** / ********* ** ********* Top ******
  • **** ******* *** ****** of ********* *********

**** ** *** ** 2 ***** ****** *** Class * ** ****'****** ********* ******.

Why Algorithm **********?

********** ****** *** ** run **** ******** *** powerful ********* **** **** hardware *** ***** ***********. Often, ***** ************ ******* have *********** ** ****, power ***********, ********, ***. that ****** ****** ******* efficiency ** ********* ***** performance.

IPVM Image

*** ********** ** ********** load, *** * ***** accuracy, *** ** * 10X ** ****. ** a ********* **** *** have ********* **********, **** typically ******* ** **** sophisticated ********** *** ******** accuracy. ** *** *** of **** *****, ** share ** ******* ** NIST ******* ** ************ real-world ****.

Video ************ ****-****, ********** *********

***** ************ ** ***** done ** ****-****, ** constricted ************, *** **** underpowered ********, ****** ********** especially *********. *** ****-**** nature ** ************ ***** processing ***** ** ****** quickly. ******* ******* ***** **** for ************ ********* ** ~15fps, **** ********** ********* at * ******** ** that (*.*., ****) ***** require * *** ** 200 ************ **********.

*** *******, ** ********* on *** *** ******, running ******** *** ** missed:

**** ** **** **** important *** *** **** 30+ *** ******** ****** to ******* ****** ********:

IPVM Image

** ****, ** *********'* accuracy ********* ** ** is******* ******* (***** *********)******* ** ** ********** frames *** ******.

Algorithm ********** *******

***** ******** ********* ********** is ******** ******* ** frames *** ****** (***) processed. * ****** ********** of *** ********* ***** testing ******** ********** ** the **** ******* **** generally ******** ***** ********** are **** *********:

IPVM Image

** ******** ********* *** robust, ********** ******* ****. However, *** ***** ********* processing ** ****, ******** resources *** ********* * critical ********** *** ********.

**** ********* ******** ******* (such ** *** ***** above), ** ******* ** determine **** ***** **********, Field ** **** *** hardware **** ****. *** example, *** ******* ***** be ***** ** ****** FoV ****** (******* ******* pictures ** *******) ** using ****-******* ******* **** multiple ****. **** ***** significantly ********* **** ***** be **** ** ***** surveillance ***** ********* *** wider **** *** ******* server *********** (*.*., ****** ********** ******* ** differing ********)

Deep ******** ************ **********

***** ****** ********** **** learning ********** *** **** accurate **** ********** *** machine ******** *******, **** generally **** * ****** computational **** ********** *** same ****** ** ****** per ******.

**** ***** ************ **** learning ********** *** ************* neural ******** (****) ******* they *** ******** ** detecting ** *********** ********/******* in ******. (*** **** on **** *** ****'****** ********* ************).

IPVM Image

** ****, *** **** of **** ********* ********** is ** *********** ******** CNN ********** ***** *********** a **** ***** ** accuracy. **** ** ******** achieved **:

  • ***** ******** ********** ** CNNs
  • ******** **** ** ****** a ******* ****** ** object *****
  • ********** *** ****

****

**** (*** **** **** Once) ********* **** ******** ** be ********* *** ******** for ***** ****** *********, are ****** **** *** open-sourced.

************** *** **** ** using *** ****** ******* for ***** **************(********* ******** *** *** class **********), ******* * two-step ******* **** *** piece, ******** ****** ************.

IPVM Image

********, **** ****** *** particularly ****** ** ***** surveillance ******* **** ***** generalized *************** ** *******. This ***** **** *** less ****** ** ******** when **** *** ********** objects ** ******* ** was *** ******* ** (e.g. *** ******* *****/******, obstructed *****).

************ ******** ************ **** **** *** using ******** ******** ** YOLO. ********** ** ********* improved ** ******** **** for **** *** ********** needed, ******* ** *** 80 ********* ** *** COCO *******.

YOLO ********/***********

****** ** ********* *** most ******** **** *********, and ***** ******* ** a *** **** ** data **********, ** ********** the ********* ** ******-*. YOLOv5 ** ************* ** deep ******** *********** ******* it *** ******** ** a ******* ******* ******* major ******** ************, ******** is ****** ** *** because ** **** ******** on ****** **** ******* instead ** ********* *** more *********** *******.

*** **** *** ********* ** ****** *** v5.

Mobile / **** ******** **** **** *********, **** ********

******/**** ****** ***** **** can ** ******** *** mobile/SoC *** ** ** modified ******** ** ********** models *** ******/*** ***. These ********** ********* ******** for ********** ***** ** important *** ******* **** learning ** ************/*** ******* with ************ ****.

****-****

****-**** ****** *** ** example ** * ******-**** more ********* ******* ** a ********** *********. ******** while******* ** * ********** ** ~*/*** ** the ****** **** ****** and ** ** ***** respective **** ****** ***** efficiency, ******** ** ***-***, increased ~** **** ******-**** and ~*** **** ******-****.

IPVM Image

***** ***** *** * large ******** ****, ***** was ** **** ****** decrease ** ************* ****, and ***** **** ** run ******* *** *** be **** ********* ** users **** *** ********* accuracy.

***********

************* ** *********** *** power ** ****** ****** neural ******* **** ****** that ******** ********** **** neural ******* ********* **** allow ****** ** ** skipped *** ********* ********* convolutions. **** ******* ************* steps **** ******** ******* steps, *********** *** *******:

IPVM Image

Reducing **********

******* *** ** ******* efficiency ** ** ****** categories. ****-****** **** *** commonly ********** *** ** default **** ****** **********, the ********* **** ** more ********* ** ******** to ***** ****** *** a ********** *** **** (e.g. ******, *******).

IPVM Image

*********** **** ******** ********** are ******* *** ********* a ****** ****** ******** (e.g. *****, ****, ******* plates). **** ****** **** to ******** ****** ******* in **** **** *** milliseconds, ***** ** ******** in ******* ************, **** facial ***********:

IPVM Image

Network ******* ***************

******* ******* *** ************ are * ******* ** analytics ********** *** ** improve **********.

******* ********** ***** ** edges **** *** ****** network, ***** ***** *** analytic *** ******. *******, the ******* ***** *** nodes *** ********* *** do *** ******* ********, but ** ********* ***** or ***** *** *******, accuracy **** ****.

IPVM Image

************ ********** *** ****** network ** ***** ***** numbers **** ******** ** smaller ******* ***************. *** rounding ** ******** ** keep **** ** *** information ** *** ******** number ***** ******** *** size, ******* ******** ****. This ********** *** ************ that *** ********, ********** the ****** ******** *** improving *****.

IPVM Image

*******, ******* *** ************ can **** ******** ******** and ******* ********** ********/*********** resources.

***** ** ** *** to ******* *** **** results, *** *****-***-***** ** typically **** ****** ** development ** ***** ****-******* or ****-************, ***** ******* in *** ********** *** memory ************, *** *********** accuracy.

Deeper **. ***** ********

****** ****** ******** *** typically **** ******** *** more *************** *********, ** developers ****** *** *** to **** **** *** if ***** ** ***** should ** ***********. ***** are *********** ******* **** both ********* ***** *** width, ******* ****** ****** or **** ******** *** typically ***********.

*******, **** ****** **** prioritize ***** **** ***** because ** ***** ** sufficient **** *** ******** it ** **** *********, however, ** ***** ** insufficient **** **** *** opposite ** ****.

********* ***** **** *********** learns ******* *********, ** each ***** ** ** increase ** *** *********** the ******** *** **.

IPVM Image

*****, ** *** ***** hand, **** *** **** to **** *********** ********* but **** **** ******** versions ** *** **** equation.

** ******** **** **** computation ** ***** **** network ******** **** **** networks ** *** ******** is ***********. **** ******** are **** **** ***** to *********** *** **** generalizable.

***** **** ******** **** more **** ** ***** and, **** ***** *** width ****** *********** ******* so ** ***** ** insufficient **** ** ** a ******* ** ******* very ****, ********* ***** will ** *** **** efficient *** ** ******** accuracy.

Combining **** ******** *** ******* ********

**** ******** *** ** combined **** ******* ******** to ******** ********** *** decrease ************* ****. *** more *********** ****** ** handled ** **** ******** while *** ****** **** is ******* ** ******* machine ********.

**** ********** **** ** ***, ***** **** ******** is **** *** ***** detection *** ********** ***** with ******* *** ** read *** **********:

IPVM Image

Hardware ************

********* ************* ***** **** learning ********** ** ******* the ********** ** *** analytics. **** * *****-********* algorithm **** ******* ****** if ** **** *** take ********* ** * chip's *********. ********** ** heavily ******** ** *** an ********* ** ***** by ********** ** *** within * ******** *********, so ************* ****** ******** to **** ********** **** full ********* ** ***** hardware.

********

*****'***************** ********** ***** *** GPUs **** *** ***** into **** ***** **** and ************, ***** ***** with ***** ****** **********. Without ******** ********** *** Intel ********, **** ** the **** ** ***** to ***** *** ********** will *** ** ***** than ******* **********.

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

********* ************ ******** ********** ** Ambarella ** ******** ****** enabled ****. ****** ********* pruning *** ************ *** converts ********* **** **** easier ** ******* ******-******** to * **** ********-******** version.

******

****** ***** ******** ******** tools** **** ********** ****** NVIDIA ********* ********** ** the ********* ** ***** choice. **** **** ********** code **** ******* ********* and ******** **** *** more ******** ****, **** translate **** ** ******-******** versions.

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

****** *** * **** version** ***** ********** ******** vision ******* **** ******* neural ******** ** ******* versions *** *** *** other ***-******* ***-*****. ********** Lite **** *** ******** for ******** ******** *** instead ********** ** ********* through ******** ***** ********* quantization.

Simultaneous ********* ******

***** ************ **** ***** involves ****-******* ****** **** lots ** ******* ** be **********, ***** ** more *********** **** ******-******* scenes. **** ********* ********** are **** ** ** run ****** *** ******* more ** *** ******* or ******. ** ** algorithm ** *** ********* enough *** *** *****, based ** *** ********* hardware *********, ** **** miss **********:

**** ** ** ***** for ****** ********* ** crowds, *** **** ***** and ******* ****:

IPVM Image

Upgrading ******** ******

********* ******** ** ********* and ** *** ****** possible. ****** **** ********** power *** **** ***** price ********* **** **** devices **** ***** ***********. It *** **** ** difficult ** ******* ********* or ******* ******* ******* overheating ****** ** ******** with ***** ***********.

******* ** ***** ***********, many **** ******** ******** in ***** ************ ***** cloud-hosted **********, *********** **-******* hardware ***********. *******, ***** analytics **** ******** ******** upload *****, ******** *** number ** ******* **** can ** *********. ********, cloud ********** ** ********* very *********, *** * monthly ********* ****, ******** to * ***-**** ******** upgrade ****.

Adding **** ******** ***** ** *******

** ******* *** ********* of ******* ******** ********* without **** **********, ****** manufacturers *** * ****** processor ********* ** **** learning **********. ****, **** Dahua *** *********, *** Hanwha ** ******* *** dedicated **** ******** ********** to ******** *********. **** allows ***** ********* ** compete **** **** ********* algorithms, ****** ** ** increased ******** ****.

** ******* ****** ********* ****, ** ***** **** deep ******** ********** **** significantly **** ******** ******* compared ** **** ******* learning **********, ***** ************* people ** ********:

IPVM Image

XNOR.ai - *** ***** ** ********** / ********* ** ********* *** ******

********* ********** ************ ** challenging, ** ** *** (not **** *********), *** involves ********* *** ****** with ***** ********* **** Apple, ******, *** ********.

* ****** ******* ** this ******.**, ***** ********* ** models *** *** ***** devices **** **** **** in **** ***-******* $** Wyze ******* *************** ***** ************* **** and ********* ****** *********.

*******,**** *** ******** *** $200 ******* ** **** by *****, *** *** ******** business ********* ****** ********* (such ** *** **** deal) *** **********. **** showcases **** *** ***** of **********, ******** $** cams ** ********** **** more ********* ******* **** more ******** ********, ***** also ************ *** ********* of ******* *** ******.

NIST ********* ********* / ****** **********

************* ********* *** * practical ********** *** ***** surveillance *********, *** ********* testing. *******, ********* ******* often ******* ********** ** allowing *** ********* ********** resources.**** ****** *********** ********* ** ******* ** algorithm *******/********** **** **** not ****** ********** ** results. ****** ****-***** ***** surveillance, **** ******* **** not ********* ******** *********, or ****** **** ***** for *** **********.

**** *********** ******* ********** that *** **** *************/*******, but ***** ********** *** prove ** ** *********** due ** ****** ****/********** for **********.

Comments (4)

**** ** *** ***** the ****** **** ** to *** ** *********** of *** ********* ** fp8 *** **** ******* GPU ********?

Agree
Disagree
Informative: 1
Unhelpful
Funny

** ** ***, * think ****** **** ** close ** *. ****** forward, ** ***** ** bigger, *** ***** *** so **** "***"...

Agree: 1
Disagree
Informative
Unhelpful
Funny

**** *********** **** ** a ******. *** ***!

Agree: 1
Disagree
Informative
Unhelpful
Funny

***** *** **** *** this ******* *******!

**** **** ******** ******* provide *** ************ *** their ********* *** **** PPF ********* *****'* ****** on *** ***** **********. But ** **** **** popular ******* ******** **** a ***** ********** ***** layer *** ** **** case *** *** ** depend ** *** ***** resolution. **** *** **** tested ****? ** **** use ******* ********* ******** for ********* *********** ** is ** **** * marketing **************?

Agree
Disagree
Informative
Unhelpful
Funny
Read this IPVM report for free.

This article is part of IPVM's 6,952 reports, 927 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