Generative AI For Video Surveillance 2024 - IronYun, NVIDIA, Verkada
Generative AI has dominated the tech news headlines in the past year, with companies like Antrophic, Meta, and OpenAI releasing various models. At ISC West, a few companies emerged as early adopters, showcasing their solutions with generative AI.
In this report, based on demos at ISC West and IPVM examination, we detail how these early adopters use generative AI and what is to come.
For background and tutorials, see:
- The Future of Video Analytics - CNNs Vs. VITs (Visual Inference Transformers)
- The Impact of Generative AI On Video Surveillance
Executive *******
***** ** ******* ********** **-***** ********* will ***** ** *** ********** ***************** in ****, **** ********* **** ***** adopters ******* *** ***** *** **** and *********** ****** ****** *** ******** already, ** ****** ******** ******** ** take ******* *****.
* ****** ****** *** ***** ******** is ***** ***** ******** ****** ** a ******* ******** ********** **** ** search ****** ********* ********. ***** *** search ***** ** ****** *** ***** to ****** *** ******** *******, ** is ******* ** *** ******** ********* from *** *******/******* (******, *******, ******** objects, ***.).
***** ******** ******* ****** *********** ****** (ViTs) ** ******* ***** ************ ** allowing ***** ** ********* ******** ******* that ****** *** ****** ** ******* images ** ********** *********. **** ******** extends *** ***** ** ********** ******* and ****** ****** ***** ********, ******** a ***** ***** ** ********* *******. However, *** ********* ************* ******* ** processing ***** ******* ****** ** ****** hardware ***** *** ******** ****** *******.
Generative ** ********** ** ***** ************
***** ********** ** *********** * ***** area ** ********** ************, ********* **** (i.e., *******) *** ***** ********** (*.*., DallE, **********), *** ***** ************, ** focus ** ********** ***** ******** ***** transformer ****** *** ******* ******** ********** with ***** ******** ****** (****).
********** ***** ******** **********, ****** ****** transformers, ******** ******/****** ******* *** **** for ******** ********, ******** ** **** commonly **** ** ***** *********. ***** transformer ****** *** ******** **** ** a ****** ** *** ** *****/***** for ***** ***** *** ******.
******* ******** ********** **** **** ****** users ** ****** *** ******* ***** a ******* ********** *** ******* ******** errors. **** **** ***** ***** ** search *** ********** ******* (*.*., "**** Musk ***" ***** ****** "***** ******").
ISC **** ***** ********
** *** ****, ******* ********* **** demoing ***** ********** **-***** *********. ** talked **** ***** ********* ***** ***** solutions. **** *** ****** ** ************ order.
******* ***-***** ******
******* ****** *** ********* ***** *** generative **. ******* ** ***** ***, IronYun ************** ****** ***** ******* ******** **********. ******* **** ** *** ****** is ******* ** *** ******** ********* by ******, **** *** ******* ******* the ********* ***********:
****** ***** ***** ****** **** ****** ChatGPT **** ****** ******* ****** *** video **** ****** *** ********. **** “Show ** *** ******** **** ****** shirts, **** **** *****, *** * backpack ****** **** *** ******* ******* between *-*** **** *** **** ****” – *** ****** **** ****** ******** search ******* ****** ***************.
****** *** "****" ****, *** ******* generative ****** *** ******, *** ********* matching ****** ** *** ******** *******. For *******, **** ** ******** "*******," several ******** ****** ****** ** ** the *******:
***** **** ****** ********* ******** (******, vehicle, *******, ***.) ** ****** ***** natural ******** *******. *** *******, ****** the ******* ****, ********* *** "* dog" ******** *** ********* *******:
******* **** ** **** **** ********* *-******* ********* ***** ******* ***** *** ******** ***** ************ scenarios *** ******** ************* **********. ******* said **** ********** *** *** ***** by ******** *** ********** ** ****** search ***** **** *** ***** ************. Even **** ********** ***********,******* ******, *** ********** ** ******** ********* the ********** **** ********* ** ***** to **,***.
****** ***** ******* ******’* ****** ****** vocabulary **** ********* ** ** ********** 50,000 *****.
*** ******* *********** ****** ***** ** ** **** and **** ** ********* ** ******, but ** **** *** ******* * date.
******* ****** *********** ******
******* **** ** **** **** **** also ********* * ****** ******** ******** where ***** *** *** ******* ******** events (*.*., “**** ****” ** “****** shoplifting”) ** ******* ******. ******* **** only *** ****** ******* *** ********** alerts ** *** ***** *******.
******* **** **** ******** ***** *** video ** * ********** *****, ***** continuous ******* *** ********* ** ******** alerts. *** *******, ** * ****** picks ** ****** **** * *****, an ***** ** *** *********, *** if *** ****** *********** **** *** item ** ***** ***, *** *********** model ********* * “****** ***********” *****.
******* **** ******* ****** "*** *****" alerts ** *** *******:
** **** ** ******* **** ** IronYun’s ********* **** **** *** ********.
****** ********** ****** ***********
****** *** ********** *** ******* ******** in *** ********** ******** *** ********** AI ** *** ****.****** ****** ******* ***** (***)**** *** * ******* ***** ***** can *** ********* ***** *** *****, which ** **** ********* **** * vision *********** ** ***** ** *** user.
*** **** *** ** ****** ****** Orin *****, ********** *** ***** ***** 1.5 *******.
******* ** *******'* *** ** ****, NVIDIA ****** *** ***** *** *** a ******** ***** ** ******* ******. In *** **** *****, ****** **** "is ***** * ****" ****** ** search ****** *** *****, **** ****** inference ****** ******** "****" ** "*****:"
***** ****** **** *** **** ********* solutions ********, ************* *** ******** **** toolkit ** ******* ********** ** ************ for ***** ************ *******.****** ******* ******** ********* ******* *** ***********:
******* ***-***** ****** **** ****** ************
**** *** *** *******, ******* ****** their ***-***** ******, ***** ***** *** input ******* ******** ******* ** ****** for ****** ** ******* ******. ******* told ** **** ***** *** ****** is ******* ** ****** *** ********, it ** *** *********** ** ******** metadata (**********, ****, *****, ***.). ******* examples ******** ****** **** ******** *******, like * "****** ******** * ****** and * ****** ***" ***** *****:
*** ***** ********, ******* ****, **** first ******** * *****-**** ***** (*********) based ** ******/******* ********* ******* *** cameras, ***** *** **** ******** ** the *****. ******* **** ***** ********* thumbnails *** **** ********* **** * transformer *** ****** ** * ******. Therefore, **** *** ***** ****** **** a ******* ******** *****, *** ******** hyperzoom ******* *** ***** ** *** user.
*** *******, **** ** ******** *** "white **** **** *** ** ***** 4," *** ******* ******** ***** ****** from ***** *.
*******’* ****** ******** ** ********** ******/******* analytics **** **** *** **** ********** thumbnails **** ************ ****** **** *** expectations *** ***** ********** ** ********. For ****, ****** ****** ** ***** ********* - CNNs **. **** (****** ********* ************).
** **** ** ******* *******’* *** search **** ** ** ********.
Contrast ** *** ***** ******
** ******** ** *** **********, **** manufacturers, **** ***** *** ***** ******, use ******* ******** ******* *** ******** search. *******, ***** **** * ******* language ******** ****** *** ****** ** certain ********** **** ** ****** **** (person, *******) *** ****/****.
** ****’* ***** *** ***** ****** testing, ** ***** ****** **** ********* for ********, **** ******* ********** **********, and ****** ******** **** ***** **** spelling ******. ** ****** **** ** improve ****** ** ********* ****** ********** to ******** ****** *** ******** * broader **********.
*******
***** ** ** *** ****** ********** AI ******** ** ***** ****** *** video ************ *** ******* *****, *** early ******** **** ***** ******' ************ and *** *****. *******, ******* ************* costs **** ***** *** ********* ********, which **** ******** ** ***** ********** become **** *********.
** ****,
****** *** *** *******. ** **** on ******* ***** ********* **** **** are *********.
** ***** ** *********** ** **** what **** ** ********** *** *** restrictions *** ****** *** ********** '*****' of ******** ********. *** *******, ******* this, *** ***** *** *** **** calculator ** *** * *** ****** that ***** **** *** *** ****** and *** *** ** ***** *** speed ** *** ****** ** ********.