Hikvision Exec On Deep Learning At IFSEC 2017

JH
John Honovich
Jun 22, 2017
IPVM

Video embedded - first half is on deep learning, second half is general products:

(2)
(1)
Avatar
Bobby Mancia Jr.
Jun 23, 2017
MIZELA CORPORATION • IPVMU Certified

In view of global terrorist attack, it’s a must responsibility of every surveillance integrator to become more involved in a premeditated vision to recognize, analyze and prepare people to prevent such event to happen. Deep learning could be the answer to this problem.

Do we agree?

(1)
JH
John Honovich
Jun 23, 2017
IPVM

global terrorist attack, it’s a must responsibility of every surveillance integrator to become more involved in a premeditated vision

Sure, but the question is whether deep learning will materially impact / reduce terrorism. It might but so far, most of the early demos / displays (not just Hikvision) have been gender / age / emotion recognition focused which has little to do with security.

U
Undisclosed #1
Jun 23, 2017

TVI with 4K and H.265

I am impressed

UM
Undisclosed Manufacturer #3
Jun 26, 2017

And POC tooAnd POC too

Avatar
Bobby Mancia Jr.
Jun 23, 2017
MIZELA CORPORATION • IPVMU Certified

I wish some manufacturers would explore more on deep learning providing the system a capability to download in the camera data base the profile of Most Wanted persons facial and voice recognition then using AI to analyze and match the suspect.

JH
John Honovich
Jun 23, 2017
IPVM

Companies, for many years, have been aware of the benefits of being able to do so. The problem is being able to do this accurately.

I am leaving voice out since surveillance cameras are not typically built or deployed to capture voice / sound at high fidelity.

But even with video and facial images, the problem is how accurate the matches of the top 10 most wanted individuals would be the vast number of surveillance cameras in the field. Surely, it's getting better but the amount of false matches and the requirements for having the authorities immediately review each one (in the hopes that there is a match) would literally make it a needle in a haystack problem.

There are variants of this that are more realistic. Let's say an organization has a concern of an ex-employee coming back to the office. In such a scenario, you are dealing with far fewer cameras and with lots of probably recent images from similar conditions. On the other hand, the ex-employee just needs to look down or tug a cap over his head to avoid such things so...

Avatar
Brian Karas
Jun 23, 2017
IPVM

I think part of the problem with this is that security cameras are often positioned poorly to get good, reliable, face shots to do that level of matching.  Sure, it can be done with some subset of cameras, but from a sales perspective it is hard to sell customers on something that only works "randomly".

Depending on how the deep learning is implemented to train the system, you might also need multiple views/angles of the suspect in order for the system to "learn" them enough for a match.

(1)
JH
John Honovich
Jun 23, 2017
IPVM

Sure, it can be done with some subset of cameras, but from a sales perspective it is hard to sell customers on something that only works "randomly".

That or deploying a whole new set of cameras specifically for the purpose, which significantly drives up cost. Some organizations do that but it typically restricts it to very high value applications.

UM
Undisclosed Manufacturer #2
Jun 24, 2017

Agree with you Brian if you focus just on video in isolation. I'd suggest that the most effective "deep learning" approaches in the future to solve this sort of problem will more heavily lean on other data in combination with video. A good example (starting to happen now) is detection and tracking of mobile devices such as smartphones using WiFi or bluetooth sensing, and then combining this with video imagery of a subject of interest at the same time. This identifies and resolves an individual's mobile device within a fairly tight location which can be linked to nearby cameras. It then becomes possible to start to associate a mobile device with some faces in local scenes - and if nothing else gives a possible list of devices that may be associated with a person of interest.

Avatar
Bobby Mancia Jr.
Jun 24, 2017
MIZELA CORPORATION • IPVMU Certified

I guess you are right, it will take more time to develop an algorithm using AI to come up with an efficient deep learning camera capabilities. Regarding the cost, I envision that governments would be able to justify the initial costing. Then, as the demand increases for more deep learning products, the cost will go down as we experience the economic of scale. 

New discussion

Ask questions and get answers to your physical security questions from IPVM team members and fellow subscribers.

Newest discussions