US Government Agency Highlights Analytics Failure

By Brian Karas, Published Jul 15, 2016, 09:10am EDT

Count the US government among those not satisfied with video analytics.

A US government agency held a workshop on video analytics for public safety. We spoke with Conference Chair John Garofolo [link no longer available] to understand what was discussed at the event and what government and public safety officials are looking for in video analytics.

In this report, we share their concerns, what they are looking for and how commercial analytics providers are still far from truly meeting that.

What *** ** **********/****** ****** *****

***** ********* ** ****** Safety Conference ***** **** **** are ******* *** ****-**** alerts ******:

  • ************* ******** ********* - fighting, ****** ******* ** water, ****** ******* ** train/subway ******
  • ***-*****/******* ***** ********* - detecting ***** **** ****** or **** *********
  • ********** ****** - ********* **** * ******* series ** ****** ** similar ** * ******** series ** ******

**** **** ********** **** common ******** ***** ******** by *** ******** ********, like *********, ** ********* *** not ************ ****** *** their *****.

 

 

Security ************* *** **** ** *******

**** ***** ***** ********* were **** ** ********** leaders ** ****** ****** analytics, *** ********** ***** ********* GE *** *** ************. This *** **** ********** at *****, *** *** complex ************ ** ********* "sophisticated ******" ** ****** safety, ******* ********** **** of ******* ****** ******** what *** ******** ****** "Deep ********".  

*** ******* ** **** traditional ******** ********* ********* are *** ******* ** simple ***** *********, *** not ******* ** ***** understanding *** ****** ***** and *** ******* *** moving ** *********** ** that *****. 

How ***** ********* *** **** *****

******** ********* ** * common selling ***** ** ********* systems *****. ********* ********* ***** ** acting ** * ***** burglar *****, **** ************ like ********* *** ****** regions, ******** ********* ** reduce *** ****** ** on-site ******, ** *** remote ****** **** ***********.

*** ******** ********* ******* is ******** ** ****** safety, *** *** ********** event ********* **** *** address ****** ****** ************ enough ** ** ** value. 

Limited ******* ****** ** ** ****

***** ***** *** * few ******** ***** ******** developers ********** ** ** what **** ** ******* for, *** ** **** have **** ******** / question *****:

  • ***** ****- ****** ***** **** $100 ******* ** *******, multiple ****** ** *******, pivots **** ******** ** non-security *** **** **** to *******, *** ******* rebranding **** *** ***** the **** ** ********* and *********** **** * standards ************ ***** ******
  • *******- ******* ****** **** child *********, ******* ****** widescale *** *** ****** ****** safety ***********
  • ***- ******* ***********, ******* stability ****** ****** *** company, ** ***** ***** on *********

Big *** ***** ******

**** *** ** ********** really ***** *** **** ********** providers *** ********** ***** has * *** ***.  The ****** ***** *** been ** ***** ** intelligent ******, ***** *** limited ***** *** ****** safety ************.  ********* ********* still ******** ** ******* systems **** *** ****** false ****** ******* **** missing ******** ****** ******** in ******* **********.

 

 

 

 

 

Comments (8)

What they are asking for sounds like the sort of solutions that are most likely to be provided by those who lead in deep learning research, large companies like Google, Microsoft perhaps... and hopefully they'll make their solutions available to the rest of us. Nvidia has some great new GPU hardware that would be ideal for incorporating modern machine learning solutions into VMSs, but I think they may need to come down in price a bit. Another thing that needs to happen is for the US government not to hamper the industry with their patent system, but the chances of getting the US government to understand their incompetence in this area zero, zip, zilch...

What they are asking for sounds like the sort of solutions that are most likely to be provided by those who lead in deep learning research

Yes, I completely agree. One thing I have said multiple times over the years is that "video analytics" is too generic of a terms to apply to as many things as we do.

I think it would be beneficial for the security industry to start defining video analytics sub-functions with words or terms that better describe how the system works or what it does. Similar to how we can describe cameras in terms of resolution or physical interface, etc.

The other thing I note, is that just by "asking" for these features, sort of implies that they have an expectation that the industry can deliver, or will be able to the not-too-distant future. That is they are aware of the significant breakthroughs that have happened in machine learning recently. If they had asked for the same solutions 15 years ago, they may have made them selves look silly.

If they had asked for the same solutions 15 years ago, they may have made them selves look silly.

10 years ago at 3VR, I talked with a number of government / public safety people who wanted these types of things and even more wild.

For example, I had an intelligence agency person ask if we could alert if a guy boards a train with a package at one station and exits at another station without the package. He was seriously unimpressed when I said no.

That is they are aware of the significant breakthroughs that have happened in machine learning recently.

I think the simpler explanation is that those are simply their needs.

It sounds like there are a few tiers structured thusly:

  • VMD - VMS or camera side motion detection. Used to be considered analytics but should really be stricken from the list.
  • Basic Analytics - Crossed line detection, dropped object, PTZ tracking, etc.
  • Moderate Analytics - Avigilon/AgentVI/Bosch IVA/etc which can be run camera side and classify objects. Basic LPR (e.g. OCR) would likely fall here.
  • Advanced Analytics - Advanced LPR, retail analytics, facial recognition, etc.
  • Mythological Analytics - All the analytics that have been mentioned as of late which seem far-fetched. Examples include: Age/gender classification, predictive behavior recognition, noise analytics, etc.
  • Deep learning/cloud analytics - the particularly complex, processing heavy analytics which are still in their infancy.

Is that accurate?

I think it is a thoughtful segmentation, but I do not think it is accurate.

For example, in "Basic Analytics", there is no way "dropped object" should fall in that category. In fact, non-human/vehicle object detection should be a unique category (IMO).

I was thinking something more like:

Perimeter Protection - Classify a human, and ignore non-human objects, in a "sterile zone" style application. This implies something that is reasonably good at picking up people, and ignoring other objects, but is not concerned with things like crowd detection, or accurately classifying a group of people.

Business Intelligence - Analytics that focus on counting and path tracking. Should be able to reasonably count a large group accurately, but less concerned with object rejection (for their standard applications you would be unlikely to have to deal with animals or other common false-alarm objects).

Behavior and Gesture Recognition - Can accurately track multiple persons in a scene, and identify basic gestures like waving arms, pushing, or actions like falling down, fighting/aggressive interaction, etc.

OCR - Anything that can read text or symbols in an image. This would include LPR functions, but also things like recognizing logo's or glyphs to alert on, or ignore.

This is just a short list, not meant to be comprehensive. Where the analytics run (camera, server, cloud) is a secondary classification, and while some functions might limit them to running only in a server or cloud environment today, they may be able to run within a camera in X years, so I would not make that part of the analytics definition.

Deep learning/cloud analytics - the particularly complex, processing heavy analytics which are still in their infancy.

Deep learning, i.e.neural networks, was in its infancy 20, 30 years ago, but seems to have come of age in recent years.

Deep Learning is the state of the art in object recognition/classification, but it is far from being the solution to *all* AI problems. It has become another bandwagon for researchers who are now proposing incremental work only.

Deep learning is still doing the lower level brain-like operations, but progress in higher level machine reasoning is just a bunch of heuristics and very little theory. I can see "200" different approaches for tracking and recognizing object on a video, but none of them have a formal theoretical framework.

BTW, my background is on signal/image processing and I remember stories from professors were army people will come to them seriously asking for mind reading technology.

Read this IPVM report for free.

This article is part of IPVM's 6,817 reports, 914 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