Avigilon CEO Solution for Terror Attacks

Published Nov 20, 2015 05:00 AM

Ripped from the headlines, spurred on by the France terror attacks, Avigilon's CEO has a solution.

Inside this note we look at his aggressive analytics plan, analyzing its potential to solve attacks.

Solution

Avigilon's CEO explained:

"**** ** ** *** ** ***** 50 ** *** ***** *** ***. If *** **** * ********* ** known ******, *** *** *** ****** when **** **** ** ** ******* areas. *** *******, ** **** **** up ** ** **** **** *** think ***** ** * ****** ****, a *******, * ****** *****, **** kind ** ******** **************, *** ***** then *** ** ***** *** **** **** attention ***** ** **** ********.

** *** ****** *********, ** ****** who ***** *** **** **** ** have ******** ******** ********. *********, ****** in ***** *** ** *** **** are ***** **** ***** * ** point *. ****** *** *** ***** to ** ********* *********, *** ******* mug ******* ** ** *** ** break **** * ******** **** ***** loiter, **** ******** ******** ********."

******** *********, *** ******* ** *** *:** minute ****:

*** ******** ** ****** ****** ************* detecting ***** ******** ** *********** **** people *** ********** ******** / ***** behavior.

****

********

******** ******* ****** ***** ****** ********** nor ***** **** ********** **** ******** in ****** ***** ** *** ********** (this *********** ** **** ********** ** *** ************* ********** *********** ** and ***** ****** ** ****** ** watch **** ** ******* **** *********).

******** **** ***** ********* *** ** one's ********* *** ******** ******** '******** ******** patterns' ****** ******* ****** ******** ** place *** * ****** ** ****. More *********, ** **** ****, ***** are **** ** ****** ********* *** any ****** ** ****** ******* *** is ***** *** ******** **** ********** can ** ********** ********* ** ********* loitering.

*******, ******** ***'* *************** **** *** ************* sophistication ** * ****** ** ****** and ** ****** *** ** ***** even ******* **** ******* ******* *******. ******* it ** ********* **** ** ********** financial ******** ***** *******, *** ** is ******* **** *** **** ***** attempt ** **** *** **** ************* ****** respect **** * '********'.

Comments (88)
SP
Sean Patton
Nov 20, 2015

Traffic cops in my city "loiter" on every major intersection... does their analytic have "ignore cops" loitering analytics?

This isn't even a knock on Avigilon, just stupid analytic marketing.

(2)
(1)
(4)
Avatar
Ross Vander Klok
Nov 20, 2015
IPVMU Certified

If it was feasible his solution is great, but since the technology is just fantasy....

UM
Undisclosed Manufacturer #1
Nov 20, 2015

(4)
(2)
RD
Robert Dyk
Nov 20, 2015

I find statements like this very irresponsible. This type of propaganda devalues the credibility of every responsible manufacturer and practitioner in our industry. I'll give the LPR statement some leeway - if prior to any event, you know what rental car company was used, by what customer and then what plate # to look for... but to even (remotely) suggest at a time like this that a CCTV system of any kind could have automatically alerted to the possibility of an impending terrorist act is a fantasy that would not even survive even the most basic "proof of concept" test. Very disappointing indeed.

(15)
UM
Undisclosed Manufacturer #2
Nov 20, 2015

This is unbelievable! Using a terrorist catastrophe for marketing and to promote his company and business and implying (in an indirect way without implicitly stating it - because he KNOWS it isn't true) that he has implementable technologies in analytics that could have prevented it. This is the next generation to ambulance chasing, very underhanded and humanly despicable.

I have lost total respect for this company. He deserves every bit of the industry backlash he is going to get on this one.

(11)
(1)
(2)
KA
Konstantin Avramenko
Nov 20, 2015

In general, I agree with him. It is a future and the question is when it will become reality.

This year there was a huge progress in a facial recognition - we have got solutions that are based on the Deep Learning. You can check current tests results (more than 30 developers) of scientific versions on the Labelled Faces in the Wild dataset. They are very impressive. And our own tests of a current and a new generation FR versions confirm it. Big angles, large and dark sun glasses, now it works with it. Of course, it will take time to create commercial versions. For now, these solutions are way too heavy. But I am sure that within next year we will see them on the market and within 1-3 years the current statement of Mr. Fernandes will not be funny. It won't be like on TV shows but it will be.

In many articles and discussions here at IPVM there is a message: video surveillance needs something new, something that will add value, something that will help to compete with Asian manufacturers, something that will push need in a high end hardware and qualified installations. Biometrics, facial recognition in connection with the video analytics is probably the best way to make it.

It won't be easy and it won't be fast but anyway we need to start working with data not only with a raw video. And it will be the worst mistake to ignore it.

Disclosure - Devotee of Biometrics. Involved in the development of both video analytics and facial recognition. Do not connected with Avigilon

(9)
(1)
(2)
HD
Henry Detmold
Nov 22, 2015

Deep learning systems (convolutional neural networks) are state of the art for various vision problems (classification, detection, segmentation etc.)

These systems tend to have lengthy training times (hours, days, weeks). Furthermore a lot of work has gone into engineering to reduce this, in particular implementation of the training phase on GPUs, so it's not obvious that much further improvement can be achieved in commercial development. Unconventional architectures (e.g. neurocomputing) might be a solution but are some way off.

Consider facial verification: if enrollment of a new person ever requires retraining of the entire network then there is a big problem if that retraining takes days. So one needs incremental enrollment that maintains accuracy. The upside is that the match phase (once the system is trained) is quite fast in software and moreover can be implemented in programmable hardware if required.

There is also a more fundamental question about how much of the information used in human facial recognition is actually accessible in the image and how much is instead in the context (i.e. humans recognise the faces they expect to see in the context they are in, and in particular fail to recognise faces they expect not to see in a given place). If it should prove to be necessary to provide automatic recognisers with the same contextual information then there is systems engineering challenge which will not align well with the current security industry.

(3)
KA
Konstantin Avramenko
Nov 22, 2015

Henry, in general you are right but do not forget that everything in our life is a compromise. I suppose that you know how Baidu has got their result - the best example of something unsuitable for real life. At the same time, this Friday I have submitted a version for the gov. project test that now can process 5 fps on a regular PC. Will it provide with the best possible result? No, it won't. However the result will be significantly better than from current FR version. The vital thing is that now we have the way. Optimization is a technical task.

(2)
U
Undisclosed #3
Nov 20, 2015
IPVMU Certified

The problem with the loitering idea is a logical one, not technological one.

Even granting the following three arguable premises

  1. Most people do not loiter
  2. Most people are not criminals
  3. Most criminals loiter right before committing a crime

it does not follow that

Most loitering occurs right before a crime

If it did then it might be worth pursuing, but sadly loitering has been shown far more often to correlate with activities such as

  • Waiting for transportation
  • Smoking on break
  • Making cell phone calls
  • Pooperscooping

Anyway, the cops don't need help finding loiterers!

(3)
(1)
RD
Robert Dyk
Nov 20, 2015

Agreed, FR is in development, shows great promise and no doubt will become a commercial reality at some point in the future. However, let's use LPR as a yardstick though, and think back to its early days very nearly 20 years ago. It is only in the past couple of years or so that universal, sustainable, cost effective and reliable LPR systems have come to market. Apply this developmental timeline to FR and we are still quite a way from marketable solutions. - let's not even discuss any privacy and networking, hardware, etc. etc. issues that would need to be solved before a vast database of watchlist faces could be proliferated or made available to all the CCTV systems out in the public domain that would be upgraded to use FR.

The problem with all electronic analytics and automated detection in my view is not whether a certain type of detection is technologically possible or not, but whether it can provide actionable intelligence. In the electronic threat detection world every positive primary alert requires a secondary inspection or followup. This secondary investigation is always much more rigorous than a primary. (ex. physical bag check at the airport if the TSA x-ray operator is sees an image he doesn't like) This secondary consumes 10x or more resources/time than the primary inspection did. If the false positive rate of any system exceeds even a very small percentage, it's value is dramatically reduced by the cost and overhead involved in the resulting fruitless secondary investigations. Each time a system operator triggers a secondary inspection based on an electronic alert he needs to be very confident that there is a high probability of a positive result, otherwise we end up with the "crying wolf" scenario where a genuine threat is potentially ignored. We can tolerate false negative or missed alerts much more easily than we can tolerate false positives. If we say our detection success is 70% then all 7 in 10 alerts had better be real.

In my opinion this is where technological development in the security industry runs into problems. Too often we forget or ignore the fact that, based on our electronic detection or analytics, we are asking for LE or security intervention, bomb squad, SWAT team, etc, or even to just stop and inconvenience a member of the public. Sound this alarm too often for a false positive and the willingness to respond disappears very quickly, not to mention concerns surrounding racial profiling, probable cause, and so on.

It is only actionable intelligence that ultimately determines the value of any analytic technology. "What am I prepared to do with the analytic output of my system?"

In short, (humerous) I think every CCTV analytics developer needs spend a month in uniform, on the street and in then front of a monitor wall in a CCTV control room and think about every event, person, behavior, abandoned object, loitering individual or other anomaly they witness and see if they are prepared to call in backup or sound the alarm and trigger a security response, or even just simply ask a stranger a question. (perhaps one could get some CEOs to do the same...)

(2)
KA
Konstantin Avramenko
Nov 20, 2015

Robert, fully agree on the complexity and false alarms issue. At the same time, please note that FR can be used in many different applications and we do not need to wait when we can reliably deal with city size security projects to start employing and benefiting from FR. Sometimes it is a question to have a result (maybe not perfect but result) or have nothing.

(2)
(1)
JH
John Honovich
Nov 21, 2015
IPVM

Robert, excellent points and I wanted to highlight your comment about the practical issues that good but not good enough analytics / FR bring:

"In my opinion this is where technological development in the security industry runs into problems. Too often we forget or ignore the fact that, based on our electronic detection or analytics, we are asking for LE or security intervention, bomb squad, SWAT team, etc, or even to just stop and inconvenience a member of the public. Sound this alarm too often for a false positive and the willingness to respond disappears very quickly, not to mention concerns surrounding racial profiling, probable cause, and so on."

Even for 'easier' applications like tripwire, a manufacturer or integrator might say something like "Isn't it fantastic that each camera only generates one false alarm a day." And, yes, from a technology standpoint, it is pretty amazing that a computer, by itself, can be so accurate. But from the operator's perspective, they are typically more overwhelmed with the "So you are telling me my staff needs to respond to X number of false alarms every night? We can't handle this. It's not worth it."

(2)
Avatar
Sal Visone
Nov 20, 2015
DWG • IPVMU Certified

Ask him to demonstrate this "watch list" feature. He definitely implies that it is a working solution. Great job in calling him out on this.

(7)
U
Undisclosed #3
Nov 20, 2015
IPVMU Certified

...within 1-3 years the current statement of Mr. Fernandes will not be funny.

Yet it will still be untrue, today.

(4)
(1)
(2)
KA
Konstantin Avramenko
Nov 20, 2015

I should write more clear that it won't be funny for other manufacturers who is laughing now. Because it is possible to get results even today. There are city size projects that already try to use FR for suspects search. And with Avigilon focus and resources... I would consider their moves seriously.

(4)
(1)
U
Undisclosed #3
Nov 20, 2015
IPVMU Certified

I'm on your side Konstantin, I think that video analytic capability, and FR is increasing everyday.

On the other hand, videos portraying future technology as if it is here today don't help anyone.

Case in point: Abu Dhabi Is a Very Safe City

(1)
KA
Konstantin Avramenko
Nov 20, 2015

Agree and Hollywood makes it even worse. It is not exclusive issue for FR though.

(3)
(7)
U
Undisclosed #4
Nov 20, 2015

One minute more and I was about to believe that Avigilon invented video analytics ...

(1)
Avatar
Mark Jones
Nov 20, 2015

50 to 300 times per day? There is a guy that needs more to do.

(1)
U
Undisclosed #3
Nov 20, 2015
IPVMU Certified

50 to 300 times per day?

Though I would imagine cctv installers would be in an even higher number of video clips per day.

Who knows? Maybe he's gone Undercover Boss at Avigilon.

(1)
UI
Undisclosed Integrator #6
Nov 21, 2015

I'd PPV that!

(1)
MI
Matt Ion
Nov 21, 2015

So, how did IPVM miss that Avigilon bought up BRS's marketing department?

(3)
U
Undisclosed #5
Nov 21, 2015

Matt,

looks like you upset because Avigilon refused to make you partner

(1)
(1)
(2)
MI
Matt Ion
Nov 21, 2015

You're funny when you hide behind anonymity and speak on topics you know nothing of.

(1)
(1)
U
Undisclosed #3
Nov 22, 2015
IPVMU Certified

Canada isn't big enough for the both of you anyway. ;)

(1)
U
Undisclosed #3
Nov 21, 2015
IPVMU Certified

looks like you upset because Avigilon refused to make you partner

They already had picked the dealer for Canada...

(1)
(1)
UI
Undisclosed Integrator #7
Nov 22, 2015

I sat in my first Avigilon presentation during the week and have to say all the marketing hype surrounding the introduction of the brand (in Australia) probably heightened my expectations of what i was about to see. The presentation was done from a private box at a large local arena/stadium so we could absorb the benefits of blanketing tens of thousands of people with up to 30MP of Avigilon magic.

As expected, lots of pixels means that you can digitally zoom to your heart's content and the images that we were viewing were very impressive. We were then regaled with the plethora of onboard analytics available in each camera including the stunning(?) fact that over 500,000 discrete objects were stored in each camera enabling it to accurately identify objects such as humans, cars, trucks and goodness knows what else.

We heard all about the value of their high quality SLR-like lenses in producing these stunning images and I slowly started to understand why the cameras were so expensive.

We then were meant to be mesmerised by the speed you could wash through multiple time-synced feeds using their patented "thumbnail-like" technology and, like many in the room, came to the conclusion that we weren't looking at anything new.

I left wondering how any of this technology would actually help in the event of a disaster at this arena (i.e. a terror attack) and came to the conclusion that most of what I saw was marketing hype for the following reasons:-

  1. Avigilon's high megapixel cameras with their expensive SLR's running at a mighty 6fps may well be able to "blanket" the arena better than anybody else but their mighty 500,000 object analytics only actually run at 1080P so all they would probably be able to determine would be that there were a lot of people at the stadium.
  2. For the majority of access areas, corridors, foyers etc etc (i.e. the bread and butter of our business) I wouldn't need expensive on-camera analytics to tell me that human beings were walking around the place. At most I would probably running loitering and people counting analytics on my VMS for probably 10% of the total cameras in key areas. For facial recognition, I would be capturing tight 1080P feeds at all my choke points and passing them on to my high power back-end servers (running say....Neoface?) for true facial recognition using whatever watch list I could lay my hands on. I don't believe any arena would be able to afford the server farm required to absorb and process multiple 30MP crowd shots using any true facial recognition solution currently available.
  3. I wouldn't need any onboard camera analytics to tell me that there were cars moving around my carpark. At best, I would be running LPR capture at major choke points and the usual small array of VMS analytics to detect congestion and keep an eye on all the "crossed line" rules I would have in place.
  4. At best, I would spend the rest of my budget on some "pattern recognition" analytics to determine unusual activity in key corridors and areas of concern using something like iCetana running on a bank of XEON processors and NVIDIA GPU's.
  5. Sure, with the 30MP cameras I could certainly go back and look at what happened after the event in glorious 6fps but I would anticipate that most of my operator's time would be taken up checking all the hotspots generated by all the other VMS-bases analytics that i would be running during the event where I am doing my best to keep 80,000 people safe

I know that this is a very broad generalisation but the only thing I took away from my two hours with Avigilon was the following:-

  1. Avigilon high MP cameras give fantastic images of people sitting at a large stadium. Unfortunately the much hyped analytics can only analyse a single 1080P stream from each of these cameras. Stadiums are a small market (and typically have small budgets). Our industry is still trying to work out the best ways of utilising high MP cameras
  2. On the surface, it looks like Avigilon's end-to-end solution is designed to impress new integrators in the small to medium end of the market. Unfortunately this market currently (in our country) would be lucky to require advanced analytics for more than 10% of their installed cameras (i.e. the other 90% are happy with record on motion, crossed line etc)
  3. Avigilon's claims that their on-board analytics result in lower network bandwidth usage and lower storage costs would be difficult to justify and are becoming less relevant given the rapid decrease in costs for both storage and network infrastructure (not to mention the fact that they are continually pushing their high MP cameras?!)
  4. I don't need on-board people detection in areas where the only motion to be detected is in fact people moving about (i.e. 95% of building cameras). The same is also true for detecting vehicles (i.e. maybe 90% of traffic cameras)
  5. The "heavy lifting" required by our larger customers to thwart terror and manage disasters will come from server based VMS solutions examining high quality live feeds using server based analytics because the cost of a CPU cycle on a camera is significantly higher than an Intel CPU cycle (which is true for both SW development and the actual hardware)
  6. The Avigilon solution is a polished marketing offering that no doubt will be successful in our market. It is a shame that the marketing capability of their competitors isn't nearly as attractive.

In summary, the comments made by Avigilon's CEO only reinforce my first impressions of the Avigilon offering i.e. a large amount of hype disguising a small amount of substance.

(2)
(1)
(3)
U
Undisclosed #5
Nov 22, 2015

UD7

Are you going to be Avigilon partner?

(1)
U
Undisclosed #3
Nov 22, 2015
IPVMU Certified

Are you going to be Avigilon partner?

After carefully re-reading his post, I think UD7 would be more than happy to consider becoming an Avigilon partner as long as they

  1. Scale back Video Analytics drastically
  2. Outsource the entire Marketing dept
  3. Get a new CEO.

How did you interpret his post?

(1)
MM
Michael Miller
Nov 22, 2015

Do you have any experience with NEOface?

(1)
(1)
(1)
HD
Henry Detmold
Nov 22, 2015

Re: 5 "heavy lifting"

I think that if the lifting is indeed heavy then server CPUs might be preferred. However:

- the cost of access to decoded frames is zero on the camera (access the frame before encoding), and non-zero on servers (which must decode before access).

- cameras alter frames to improve them for human consumption; only on-camera processing can access the frame prior to alteration, and this may yield better results for automated analysis

- hybrid approaches involving image analysis on the camera and high level analysis on server are common

In the 1980s when distributed file systems started to become common it was said: "clients have cycles to burn". This wasn't so much about client CPUs being cheaper per operation but more about the client CPU having more available capacity as it was not needed for anything else, whereas server capacity was always highly in demand. A similar point obtains with on-camera processing capacity: the spare capacity is not in demand for anything else.

(2)
(1)
UI
Undisclosed Integrator #7
Nov 22, 2015

"clients have cycles to burn".

The trouble with today's clients (i.e. cameras) is there has largely been a "race to the bottom" to install the lowest cost processor on a multitude of different OS platforms. Axis have probably the most "open" environment for using onboard cycles and yet the interest has been moderate at best.

If the clients could agree to perhaps an "ONVIF compliant" development environment on a common hardware platform (which would never happen), I would agree.

- decoding could be done by onboard GPU's (e.g. Intel Quicksync)

- most analytics prefer to take their own untouched feed from the camera

- hybrid requires "openness"....need I say more?

(1)
(1)
(1)
HL
Horace Lasell
Nov 22, 2015

Undisclosed 7 Integrator indicates that "The 'heavy lifting' required by our larger customers to thwart terror and manage disasters will come from server based VMS solutions examining high quality live feeds using server based analytics because the cost of a CPU cycle on a camera is significantly higher than an Intel CPU cycle (which is true for both SW development and the actual hardware)"

It seems as if the issue might be as much communications as processing costs. Otherwise, mightn't the economic benefit of less expensive processing at the server have encouraged experimentation with alternate architectures?

It's conceivable that the video could be exported raw from the camera. This would not impose any decoding penalty at the server. If processing were more efficient at the server, that could as easily be where filtering and compression would occur, which could make analytics equally applicable at the camera or the server (because raw video would be available at either location).

Today, on the camera, application Specific Integrated Circuits (ASICs) do the heavy lifting of converting from Bayer filtered pixel values into color images and to H.264 compressed video. I wonder how costs might compare for this very specific processing, camera side vs server side?

Enterprise networks might be able to sustain the 75% load of streaming 1080p raw video at 30 fps, or 4K raw video at 6 fps, or 30MP raw video at 2 fps, to a local server across gigabit ethernet.

That would seem to require the server to have a NIC per camera, to make raw video available at the server without a decompression penalty.

It seems that Henry Detmold is spot on when he suggests that "the cost of access to decoded frames is zero on the camera (access the frame before encoding), and non-zero on servers (which must decode before access)."

Regardless of whether or not new, very compelling capabilities could justify additional costs, ... whether it is through the cost of decoding, as in the current architecture, or the cost of moving a much greater mass of raw uncompressed data, as in this alternative thought experiment, server access to decoded frames seems to bear a significant additional cost. This is without considering even more demanding raw streams, such as 30 MB at 6 fps, that might require 10 GE or even higher network link capacities.

If it's a choice between communications costs and decoding costs, then when compelling capabilities become reasonably amenable to server resources, we might just see affordable custom ASICs decoding H.264 on the fly at the server. These ASICs are probably in use today in televisions and monitors.

(1)
(1)
U
Undisclosed #3
Nov 22, 2015
IPVMU Certified

It's conceivable that the video could be exported raw from the camera. This would not impose any decoding penalty at the server.

Good news!

Uncompressed server side analytics could be developed by just using CVI/TVI/SDI/AHD. Not sure if they do this currently, but uncompressed video gets to the DVR, so why not.

Its kinda crazy to build out a enterprise TCP/IP LAN to carry raw video from a camera to a server a hundred feet away.

Just get a wire and be done with it! :)

(1)
(1)
HL
Horace Lasell
Nov 22, 2015

Sounds great! If all these pieces hang together, could it suggest that the future of analytics is with analog?

U
Undisclosed #3
Nov 23, 2015
IPVMU Certified

Not so much with Analog, as with just using the whole wire to transmit the whole signal. Cause it can fit. And since the wire is not shared with anybody anyway, why not?

HD-SDI is digital and would likely have the best signal for short runs.

Related:Are Camera-Side Analytics More Accurate?

(1)
U
Undisclosed #3
Nov 22, 2015
IPVMU Certified

Another idea is simply to do the prep work of meta data creation on the uncompressed stream and forward those results to the server.

That way you can have the best of both worlds.

HD
Henry Detmold
Nov 22, 2015

Intel have for some years sold server (Xeon) CPUs with embedded graphics, including H.264 ecode and decode, MJPEG decode and (I believe) H.265 decode in hardware. These are the E3-12xy CPUs where the digit "y" is either 4, 5 or 6. One significant limitation of these CPUs is that they are uni-processor only and also they are limited to four processor cores. So in a conventional 1RU server enclosure one would have the following choice:

  • E3-12x5 uniprocessor with video decode hardware and 4 CPU cores, or
  • E5-2630 (say) dual processor without video decode hardware but with 16 CPU cores

The latter provides substantially lower TCO per processing cycle for general computation so it will be a bit of a balancing act.

Intel are in the process of completing the acquisition of Altera who are the second largest FPGA manufacturer (it's a $16BB acquisition so it is spread over about a year). Their announced plans are around:

  • Addition of FPGAs to Xeon for the data center market
  • Addition of FPGAs to embedded processors for the IoT (and motor vehicle) market

In both cases the idea is to make it easier for their customers to move key processing from software into hardware; replacing many ASICs and ASSPs. It seems likely that the IoT (and motor vehicle) version of this will make its way into cameras, particularly since motor vehicle systems already have extensive video capability.

(1)
UM
Undisclosed Manufacturer #8
Nov 22, 2015
If you think of it as a trend, or a solution that we'll see in the future. I agree with him. Yet, you're right that this is barely possible with current technologies and solutions.
(1)
JH
John Honovich
Nov 22, 2015
IPVM

"If you think of it as a trend, or a solution that we'll see in the future."

I know we are talking a lot about face recognition in the comments, but the "abnormal movement patterns" is even optimistic to say in the future. Sure, the future is infinite but to seriously claim that reliable 'abnormal movement pattern' analytics are anywhere in the near future is groundless (sorry BRS Labs).

As for the face recognition, Avigilon's CEO grounded it in the "Most of us are on video 50 to 300 times per day" point. If you hooked up the best facial recognition in the world to those existing cameras (again, Avigilon does not have / offer face recognition at all), the results would be a disaster as overwhelmingly real world cameras are not setup for facial recognition. Konstantin is certainly implying deploying new cameras and even then, the likelihood of stopping a terror attack versus the massive amount of incorrect matches one would have to go through would be incredible.

(2)
KA
Konstantin Avramenko
Nov 22, 2015

Certainly, it is not possible to use current security cameras for FR in most cases. However, is it bad for the industry to have an opportunity to up-sale hardware, software and services?

I watched the video one more time and, despite the fact that Avigilon's CEO was trying to promote Avigilon (who won't?) and messed a little with now and future, I think it is a good strategy to focus on the intelligent solutions to add value to the expensive hardware. The bigger issue is that the host was trying to get a magic solution for preventing terrorists attacks. Avigilon's CEO answers clearly did not satisfied her "It sounds like it takes time - bye-bye". Everything takes time and clearly the "smart video surveillance" is not a magic solution but just a tool that has to be used properly to provide with value.

(1)
JH
John Honovich
Nov 22, 2015
IPVM

"despite the fact that Avigilon's CEO was trying to promote Avigilon (who won't?) and messed a little with now and future"

Genetec, for example, didn't. Many of the biggest cities in the world use their video surveillance systems, they have a sizable office in Paris not far from the attacks, etc. And if anyone had the bonafides to 'mess a little with now and future' given Genetec's experience and relationship to Paris, they could have.

"is it bad for the industry to have an opportunity to up-sale hardware, software and services?"

Yes, it is irresponsible to as you justify it, mess 'with now and future'. We are here to provide people with security solutions that work, first and foremost, not 'up-sale' people on things that do not work now.

"The bigger issue is that the host was trying to get a magic solution for preventing terrorists attacks."

The host was looking for magic but a responsible leader makes it clear what can or cannot be done and does not simply tell people what they want to hear so they can get a sale.

(1)
(2)
KA
Konstantin Avramenko
Nov 22, 2015

"is it bad for the industry to have an opportunity to up-sale hardware, software and services?"

Yes, it is irresponsible to as you justify it, mess 'with now and future'. We are here to provide people with security solutions that work, first and foremost, not 'up-sale' people on things that do not work now.

Please do not mess with my message. The point on the up sales was regarding the need to install new cameras. And FR is not something for the future only if you try to think about something else other than preventing terrorist attacks.

(2)
(1)
JH
John Honovich
Nov 22, 2015
IPVM

"The point on the up sales was regarding the need to install new cameras. And FR is not something for the future"

Avigilon does not sell facial recognition, so not sure what or why Avigilon's CEO is even selling here.

"And FR is not something for the future only if you try to think about something else other than preventing terrorist attacks."

The topic of the interview and this discussion is preventing terrorist attacks.

(1)
(3)
(1)
MM
Michael Miller
Nov 22, 2015

Genetec, for example, didn't. Many of the biggest cities in the world use their video surveillance systems, they have a sizable office in Paris not far from the attacks, etc. And if anyone had the bonafides to 'mess a little with now and future' given Genetec's experience and relationship to Paris, they could have.

John where is your outrage for this video that Genetec is promoting?

(2)
(1)
JH
John Honovich
Nov 22, 2015
IPVM
MM
Michael Miller
Nov 22, 2015

LOL... John "CCTV COP" Honovich

(2)
(2)
JH
John Honovich
Nov 22, 2015
IPVM

Genetec has deleted the tweet.

That's a good example for Avigilon's CEO here. It's a mistake, recognize it and move on.

(2)
U
Undisclosed #3
Nov 22, 2015
IPVMU Certified

LOL... John "CCTV COP" Honovich

Sometimes the offense forces one to take matters into their own hands.

Call it Avigilante Justice.

(3)
KA
Konstantin Avramenko
Nov 24, 2015

As I said John, business is business. So, there is no need to overreact on a such kind of things. Of course, you can write that it is a different thing and there is no connection and etc. But maybe you can then also explain how VMS can be AntiTerrorism surveillance?

Hope that Avigilon, Genetec and other companies continue to innovate to provide us with more ways to deal efficiently with modern threats including terrorism. Let's focus on the technologies and solutions not on the "who say what" and other gossips.

(2)
JH
John Honovich
Nov 24, 2015
IPVM

"But maybe you can then also explain how VMS can be AntiTerrorism surveillance?"

First of all, that's simply wrong. A VMS alone was not presented as 'antiterrorism'.

As the actual post explains in line one: "Microsoft presented information on its “intelligence-led, first response” campaign" And then in the cited Microsoft post, Microsoft notes "The team decided to highlight our Microsoft Global Security Operations Center (GSOC) case study and our “intelligence-led, first response” (ILFR) campaign; a global best practice for anti-terrorism efforts and includes our Microsoft solutions." Genetec is simply one component of that.

And it's a scheduled presentation at conference, not speaking directly to a reporter about a specific terror attack that happened days before.

Konstantin, in nearly IPVM post about facial recognition, you defend it but you irresponsibly conflate what Avigilon's CEO said he could do now with face recognition and abnormal behavior with their own technology to what you hopefully, in the future could do.

(1)
KA
Konstantin Avramenko
Nov 24, 2015

John, I do not remember the moment when I said that Avigilon can do it now but anyway I propose to stop the discussion of "words". I do believe that you understood my point and it will be waste of time to continue this discussion.

(1)
JH
John Honovich
Nov 24, 2015
IPVM

"John, I do not remember the moment when I said that Avigilon can do it now"

I know you do not believe that Avigilon can do it now. And thank you for making that clear because that is my core issue - don't make specific claim about things (like he did) that may not work until some unclear point in the (potentially distant) future (like abnormal movement patterns).

MM
Michael Miller
Nov 24, 2015

Let's focus on the technologies and solutions not on the "who say what" and other gossips.

Oh wouldn't that be nice but John is way more interested in twisting everyone's words around for his benefit.

(4)
JH
John Honovich
Nov 24, 2015
IPVM

Mike, when talking about 'benefits' the real benefit is for you defending your #1 business partner, which is your main use of IPVM comments.

(2)
(1)
MM
Michael Miller
Nov 24, 2015

I think a thank you is in order. Out of 10K users only 50 users post regularly. I am far more interested in talking tech then this gossip crap you post to generate clicks.

(1)
(2)
JH
John Honovich
Nov 24, 2015
IPVM

Mike, I am happy to cancel your membership and refund the existing payment.

We both know IPVM is your #1 marketing vehicle. Fight with IPVM to get attention for yourself and approval from your business partner.

(1)
(1)
(1)
MM
Michael Miller
Nov 24, 2015

John please stop this tit for tat verbal attack. The perception that you have (and project) of me versus the reality is completely false and wrong.

I am interested in the technical info and if I see false or misleading info I will call it out for any for any manufacture. I have more important things to work on then defending myself from your false claims and you have more important things to do then make them up. Move on and have a good day.

(3)
(1)
UI
Undisclosed Integrator #6
Nov 24, 2015

(1)
(2)
U
Undisclosed #3
Nov 24, 2015
IPVMU Certified

if I see false or misleading info I will call it out for any for any manufacture.

Do you mind linking to a post where you called out Avigilon for misleading info?

(3)
(1)
(2)
UM
Undisclosed Manufacturer #1
Nov 24, 2015

(1)
AT
Andrew Thomas
Nov 22, 2015

Facebook facial tagging is already being used by law - enforcement. I could see where Facebook creates a real-time API allowing video security systems to submit facial detail in much the way LPR systems retrieve data. Perhaps the first place where this might be feasible is in points of entry like airports, border crossings, etc. I've been using Picasa for a long time for my family photos, and it's amazing how often it finds a friend or kid from older poorly focused images.

Long ago (mid-late 80's), a friend that owned a string of 1 hour photo shops and I pontificated that digital cameras will never NEVER 35mm film.

For us and our clients, we promote analytics as a proactive tool to prevent or alert on potential crimes. Real example in progress: behind our building, our analytic cameras have broken the area up into many zones... (loitering near my RV) West entry virtual fence, East Entry virtual fence , Near Overhead doors, etc...

The analytic events trigger a PTZ mount with a Military High Intensity white LED, to move to the preset and place this light in the general area of the detection. (overnight hours), and have an outdoor horn blast for 5 seconds, On the front of the building a blue light begins to flash. We are very close with the police department, and they know the blue light means someone was detected behind our building in the last 5 minutes.

As the perpetrators move about, the light moves from zone to zone. Also on that mount is a very good 2MP with an extremely long lens. that goes to focus presets in the process.

that's the extent of how we are using and promoting Analytics, and I'll never say what something won't do in the future.

analytic rear

(1)
(1)
JH
John Honovich
Nov 22, 2015
IPVM

"I'll never say what something won't do in the future."

Anything can happen in the future but Avigilon's CEO was not speculating about the far future.

"I've been using Picasa for a long time for my family photos, and it's amazing how often it finds a friend or kid from older poorly focused images."

That's because it limits itself to a relatively small number of people, which makes the problem far less difficult and far less likely to be wrong. Face recognition in public areas does not have that luxury. You need to match against tens or hundreds of thousands of faces a day and cannot reject a match simply because they are not in the group of one's friends or families.

(1)
AT
Andrew Thomas
Nov 22, 2015

Agreed on all points, and the technology is in it's infancy. FingerPrint and DNA databases are extremely fast with the search results, (not defending accuracy) and facial technology is extremely good when using passport photos. But who looks like their passport photo? :)

(1)
Avatar
keith maxwell
Nov 22, 2015
Northeast Remote Surveillance and alarm, LLC

That's a typical CEO answer totally out of touch with reality. I don't have experience with the corporate echelon of the security world.

I have sat sales meetings with Diversey Lever , Johnson and Johnson during New product roll outs and the sales force said to each other wtf are we gonna do with that no one wants that

It's a very common corporate mentality to think you know everything when you don't. It could have something to do with spineless middle management telling the head honcho what he wants to hear. I have no idea in this case.

However it's a fabulous idea he had maybe he should work in the field for a few days so he can enjoy reality like the rest of us.

(1)
(1)
(1)
U
Undisclosed #3
Nov 22, 2015
IPVMU Certified

It could have something to do with spineless middle management telling the head honcho what he wants to hear...

No, I think that's been fixed...

(1)
(1)
AT
Andrew Thomas
Nov 22, 2015

Fortunately, the members of this forum, and people in the know are able to separate fact from fiction, and ill timed, or poor exploitation of events. Unfortunately, political leaders on a much larger stage are doing the same thing.

(1)
(1)
Avatar
keith maxwell
Nov 22, 2015
Northeast Remote Surveillance and alarm, LLC

In any event he lost all credibility.

(1)
(1)
(1)
(1)
AT
Andrew Thomas
Nov 22, 2015

(1)
FY
Frank Yeh
Nov 23, 2015

Marketing hype like this is one reason the video analytics market has not taken off. Setting unrealistic expectations leads to disappointing results and unhappy customers.

Video analytics are getting better with deep learning, but the training requirements are a big deal and will continue to be for quite some time. Humans' cognitive capabilities are developed over years when we are children and this takes years. This being the case, the best approach is to position video analytics as a tool to assist humans, not providing final answers but reducing the time it takes for a human to get the answers.

When I try to explain this to customers I use an analogy of internet search tools. EG when you search for something on the internet using Google, Bing, Yahoo, etc, do you always get the single best answer for your search? Not likely! So would you be better off finding things on the internet without one of these search engines? Definitely not! Video analytics are like search engines for video_ they will help a lot but you still have to make the final decision.

(3)
KA
Konstantin Avramenko
Nov 23, 2015

Good analogy

Avatar
Ethan Ace
Nov 23, 2015

Agreed, humans do have to make final decisions, but lets extend out your search engine analogy with real video analytic examples.

  • Search Google for "vehicle crossing line on road" and instead it gives you results for a bunch of leaves blowing in the wind.
  • Search Bing for "human enters area" and instead it shows you results for a streetlight.
  • Search Yahoo for "human crossing line" and you get a bird.

Would you expect that users would continue to use search engines with results of that quality? Those are all actual results from real world tests, btw, not exaggerations.

(1)
(1)
KA
Konstantin Avramenko
Nov 23, 2015

Ethan, try to make these searches in Google Images. Are you ready to stop using Google?

(1)
(1)
FY
Frank Yeh
Nov 23, 2015

Not sure what real world tests with whose products you are referring to but ours does a pretty good job for all of these use cases_ not perfect but good enough to be of value.

One thing we have noticed more often than not is that camera positioning and selection are not done with analytics in mind. Integrators who understand how analytics work would be of huge value to the industry, but the typical mindset is to provide images that are "pleasing" to the human eye, which very often are inferior for analytics. A recent example is a customer who is using analytics for perimeter protection of a very long fence line. When complaints about the analytics came to us, we discovered that the cameras were positioned so that about 50% of the field of view was sky. Nice pictures if you are shooting a travel video, but if you want to know if someone is approaching a fence, you're better off not wasting pixels on the sky. Also, since half of the cameras were facing either East or West, they were getting very extreme lighting effects in the early morning and evening. Repositioning the cameras so the horizon was just above the top of the images improved analytic performance and also increased the total area covered by the cameras.

(1)
(1)
HD
Henry Detmold
Nov 24, 2015

Ethan's examples all have a somewhat unusual property---indefiniteness---that distinguishes them from inputs to a human visual system and also from the input to a surveillance visual (video) system.

Suppose one adds the indefinite articles (underlined) into the queries:

  • A vehicle crossing a line on a road
  • A human enters an area
  • A human crossing a line

A human interpreting the situation described by the first query (for example) is anchored in various ways:

  • The human's viewpoint must by definition include a road (else the human won't be interpreting a scene containing a road) and in that case it must be a particular road (the road the human can see). The human most probably knows which road it is too (could be mistaken, of course).
  • Similarly there must be a particular line; name the line on the road.
  • Only the vehicle component is indefinite for the human (the human will do some kind of "classification" to determine that the thing seen is in fact a vehicle).

So for the human the query is (much simpler):

  • A vehicle crossing that line on that road

I.e. there is a lot of structure implicit in the question (query) that simplifies the problem of interpreting query results. So to provide an equivalently "easy" query to an automatic system one would have to provide the prior context, so the query becomes:

  • Here is video of a scene containing a road with a line on it. Show me any vehicle that crosses the line.

It is obvious that there is a much greater chance that an automatic system would produce sensible results when given this query , and equally obvious that this query is a much less powerful (and in principle less useful) than the original.

One can also note that the establishment of the prior context is quite similar to the setup of "rules" like tripwires etc in alerting analytics. Perhaps providing this kind of structure is important for search as well? Setting up rules is rightly regarded as tedious and time consuming and this perception would be even stronger if the purpose of the rules was simply to enable search (i.e. no-one would do it). So one ends up wanting to have the rules / structure somehow learned by the system...

(1)
RD
Robert Dyk
Nov 23, 2015

To add to Ethan's point a little,

Video analytics are meant (and sold) to report on a very precise visual condition observed in real time, not query a vast database with very complex algorithms in order to return hundreds of possible results.

In the search engine analogy, how often is the first hit on the first page returned by Google the ideal answer to your query? A spectacular result would be let's say 25% of the time. Imagine if video analytics were actionable only 25% of the time! This would be completely unacceptable and would be rejected by even the most patient CCTV operator.

The fundamental difference in my view is that the human element in the Search Engine example takes over after the search when the returned results are evaluated by the operator. As surveillance specialists who will eventually rely on analytics routinely, we need the human factor to be emphasized before the search with very tight definitions, and image processing that returns a clear binary result. - "ALERT" or no alert.

The (desired) analogy to an internet search might be "has this URL been updated today?" This would return a very clear yes or no. This is the reliability that is ultimately required of video analytics.

(1)
(1)
KA
Konstantin Avramenko
Nov 23, 2015

Robert, I believe that the point of analogy is that despite the fact that you cannot get the reliable/accurate results all the time it is still worth to use it.

Yes/No approach includes too many ifs: if customer has good cameras, if cameras were installed properly, if suitable rules were applied correctly and so on. We do not have AI yet to forget about human factor and believe 100% in the result of analysis.

(1)
FY
Frank Yeh
Nov 23, 2015

Robert,

You're referring to real-time alerting, which is one of the two main value propositions for analytics. IMHO analytics are just as valuable for forensic searching of video (and they tend to work better for this as it's much easier for people to ignore incorrect results when searching than when receiving alerts).

When used for forensics, using them is very similar to using a search engine. You use them to search for video that is of possible interest from thousands of hours of recorded video. After the search results are returned, a human then evaluates them.

(2)
JH
John Honovich
Nov 23, 2015
IPVM

"IMHO analytics are just as valuable for forensic searching of video (and they tend to work better for this as it's much easier for people to ignore incorrect results when searching than when receiving alerts)."

Your opinion is in the minority of security professionals. From speaking to security directors and buyers over the years, give them a pick between alerting and search, and 9 / 10 easy take alerting.

Examples:

Alerting - alert me that the terrorist is on site so I can stop him.

Searching - search the video and show me where the terrorist went as he killed all those people so I can understand how he killed all those people.

Alerting clearly has higher value in security. That said, I agree with you that searching is easier to do but it's no surprise that sales people like Avigilon's CEO leads with alerting.

(1)
(1)
FY
Frank Yeh
Nov 24, 2015

Hi John, strange_ your comment does not show up in the web page but I see it in the email notice I got.

I agree that 9/10 customers are all about the alerts at first but that's really somewhat misguided (again, IMHO). Yes, it's sold this way and people have this perception that this is what they want but too often reality does not meet expectations. Even when alerts perform well, people really don't realize what they are going to get until things get turned on.

Take the example of loitering that has been discussed in this thread. If you generate accurate alerts any time someone is stopping in an area of interest, what will typically happen is the user will be educated about what really happens in the area and will then realize that they really don't want an alert every time someone stops there as a vast majority of the alerts will be alerting them to benign activity.

Sure there are some use cases where the activity of interest is almost always cause for alarm, but these are more uncommon than most people realize.

Another reason that alerts fail to live up to expectations is that much of the time you really don't know what the activity of interest is before it happens. It's not until an incident happens that you know what you are looking for.

Perhaps I should have said that forensic searching delivers as much or more value than real-time alerting.

Another example is what happened with the Boston Marathon bombing. Nobody has a good enough abandoned bag alert to have caught every bag being left in the crowd at that event, and even for those that would have been detected, how many people do you think left something on the ground at that event? Probably too many to have made those alerts valuable. However, once the incident did happen, authorities were able to use video to search in the area of the explosion and did observe the bag drop. At that point how valuable would it have been to have analytics assist them in searching all of the video they had to search through to find people matching the description of the bag dropper? My understanding is that this search was performed through sheer manual effort.

Your description of what people want, "Alert me that a terrorist is on site so I can stop him" is not within the realm of realistic possibility for too many reasons to discuss here. Your description of what search would be used for misses a lot of the value of search. EG in Paris, tracking down the terrorists that were not killed or apprehended has been a big problem. How about being able to search through all available video to track not just where they went after the incidents but where they came from prior to them? That is how a lot of perps are ultimately tracked down_ they're usually more concerned with covering their getaway than concealing where they are before an incident.

JH
John Honovich
Nov 24, 2015
IPVM

"Your description of what people want, "Alert me that a terrorist is on site so I can stop him" is not within the realm of realistic possibility for too many reasons to discuss here."

Yes, I agree, ergo my commentary in the original post :)

My point was not to say that alerting works but to say that people value alerting more than search.

"Perhaps I should have said that forensic searching delivers as much or more value than real-time alerting."

Yes, I agree. That's unfortunately for many end users an disappointing scenario, given the higher value / hopes for alerting.

"Your description of what search would be used for misses a lot of the value of search. EG in Paris, tracking down the terrorists that were not killed or apprehended has been a big problem."

That's also an extremely hard problem to do with video analytics, as you well know. It's one thing to say show me all the people in red jackets in front of the post office yesterday. It's another to use video analytics to track where that person in the red jacket went to today.

I am also not saying this should not be tried but it's an expensive, difficult proposition even for the handful of super high risk security users that would attempt it.

FY
Frank Yeh
Nov 24, 2015

"That's also an extremely hard problem to do with video analytics, as you well know."

Yes, but it's harder without video analytics.

"It's one thing to say show me all the people in red jackets in front of the post office yesterday. It's another to use video analytics to track where that person in the red jacket went to today."

Fair enough, the larger the temporal discontiguity between events the harder it is to correlate them. However, this is not what I was suggesting. The concept of tracking a person across multiple cameras is fairly common nowadays, what I was describing is back-tracking a person or vehicle across multiple cameras to discover where they started, what house they came out of before they got in the vehicle that they eventually drove past a restaurant while spraying gunfire into a crowd. If you can come up with an address, you can then apply information and identity analytics to understand who these people are, who they network with, etc. So the video analytics is just one piece of a larger puzzle that includes humans and other analytics systems, which brings me to your next comment...

"I am also not saying this should not be tried but it's an expensive, difficult proposition even for the handful of super high risk security users that would attempt it."

Therein lies the rub. Is there sufficient ROI to justify the expense? This is a dilemma that plagues the security industry as it's extremely hard to quantify the value of good security. Return on Investment for security systems can't be measured in positives, it's more an avoidance of negatives. So how do you measure the cost of 100 lives? The cost to society of basically shutting everything down for days or weeks? I've seen some studies and they are reasonably compelling but do not go far enough. Even given that kind of information, how do you communicate it with customers so they understand that the costs are justified? This may be a harder than video analytics!;)

(1)
(1)
MM
Michael Miller
Nov 23, 2015

Frank I agree and we are seeing customers get very interested when we show how to search using analytics. Proactive analytics are not a good fit for all customers as most customers don't proactively monitor their systems. Almost every customer uses the system to search and review video and using analytics can reduce that search time to seconds and minutes.

(1)
(1)
JL
Joshua Letourneau
Nov 24, 2015

Just playing Devil's Advocate here, but what is the alternative? Simply lay down and not attempt to innovate (and share intel) around a growing and urgent worldwide threat? At this point, Video analytics are not evolved enough, not reliable enough, etc. However, when combined with other agency data (city/state/federal), they can assist in mitigation and response. My comments here are less about Avigilon and more about the need for a) improvement and additional innovation in the Video Analytics space, and b) better sharing of intel among all global stakeholders. As we all know, there are tremendous challenges in these arenas ... but doing nothing is not an option. Al Qaeda and ISIS/ISIL are innovating and planning right now, as we speak. We need to be doing the same.

UI
Undisclosed Integrator #6
Nov 24, 2015

Simply lay down and not attempt to innovate (and share intel) around a growing and urgent worldwide threat?

I don't see anything saying we shouldn't continue to innovate... only that innovation shouldn't be portrayed as doing more than it actually can. That just raises unrealistic expectations, leading to disappointment, and eventually less tendency to accept and implement real innovation when it does come along.

(2)
UM
Undisclosed Manufacturer #9
Nov 24, 2015

I would recommend this guy to sell hot dogs...

UM
Undisclosed Manufacturer #10
Feb 25, 2016

Anyone that has "real-world" experience with "intelligent" video analytics understands it is nothing more than a "tool" with limited accuracy and results. While sales people would like to say otherwise there are just too many factors both logistical and environmental that render them limited in their usefulness. Any face, pattern or other recognition program has specific requirements to even have it minimally function. One of the challenges is a face recognition analytic as an example requires a specific number of pixels on target and a very flat angle. When used in a subject-compliant environment indoors it has some accuracy. When attempting faces in the crowd outdoors in changing sunlight good luck!

KA
Konstantin Avramenko
Feb 25, 2016

One of the challenges is a face recognition analytic as an example requires a specific number of pixels on target and a very flat angle. When used in a subject-compliant environment indoors it has some accuracy. When attempting faces in the crowd outdoors in changing sunlight good luck!

To recognize something you need to see it. Why for LPR it is reasonable to expect requirements to image quality but in case of FR it's unacceptable? I wrote it previously, face features are not black lines on a white background and obviously we need to have more pixels to be able to process image.

Make a search on the development in artificial neural networks. Deep Learning facial recognition shows much better results and more robust to the angles and complex light.

From the latest test in the city environment, I, personally, have got feeling that only a highly trained pros can search better but it is not impossible even for them to stay alert for hours.

It will be a huge mistake to reject FR.