Dahua Claims Their Deep Learning Can Determine Whether Drivers Wearing Seat Belts

JH
John Honovich
Jul 20, 2017
IPVM

Lot of buzz words in this video and quite a number of big claims including that they can identify whether drivers and passengers are wearing seat belts. As they say, 'big if true' but seems ludicrous to be able to reliably be able to detect something that nuanced, especially on moving vehicles, with tints, etc.

In fairness, this video is lot better than their previous analytics marketing video: This Dahua Safe City Marketing Video Is Irresponsible

Avatar
Brian Karas
Jul 20, 2017
IPVM

This is a common request from some markets, I recall it being asked several times at VideoIQ. 

Seatbelt detection was mentioned in a previous Hikvision/Movidius press release.

This company in Dubai claims to be able to detect seatbelts (and other traffic violations).

A whitepaper from 2013 describes a method for doing seatbelt detection.

Kria has a patent for seatbelt detection, though it is not listed in their current traffic monitoring system.

Another whitepaper from 2014 on the topic.

As usual, Dahua is playing catchup with this.

JH
John Honovich
Jul 20, 2017
IPVM

Catchup on making it work or catchup on hyping something that does not work?

U
Undisclosed #1
Jul 20, 2017

I would think they would focus on deep learning how to not get hacked constantly. Some of these companies are delusional in their approach. 

(1)
(2)
U
Undisclosed #2
Jul 20, 2017
IPVMU Certified

I think it's entirely possible to determine with high confidence that someone is wearing a seat belt thru video analytics.

On the other hand determining with high confidence that someone is NOT wearing a seat belt is bound to have false positives.  

Therefore since the likely implementation would primarily be interested in the 'NOT' case, I'm doubtful...

(1)
JH
John Honovich
Jul 21, 2017
IPVM

I think it's entirely possible to determine with high confidence that someone is wearing a seat belt thru video analytics.

Even in that case, the analytic may have low false positives (i.e., identifying someone as wearing a seatbelt who is not). However, because of obstructed video frequently, it would still miss lots of people wearing a seat belt.

That said, I agree with your point about the implementation being about focusing on those NOT wearing a seat belt. 

For people who hypothetically what to know about who wears vs does not wear seat belts, is this for statistics / trend analysis or real-time alerting / intervention? The statistics are going to be way way off because of the issues described above. You might get some correct alerts but is the goal to mail them tickets or to dispatch a police officer? I am genuinely curious if there is a real demand for this outside of it being something neat to do.

Avatar
Brian Karas
Jul 21, 2017
IPVM

The times I had heard it coming up, it was for active traffic enforcement, alert police to pull over/ticket drivers not wearing seat belts.

This was mostly in Asia and UAE.

 

U
Undisclosed #2
Jul 21, 2017
IPVMU Certified

Slightly(!) easier, HOV abuse detection thru video analytics has recently made great strides, though these seem like fairly expensive systems.

Even then they are touting only 98.9% accuracy for 2-person HOV capture, as in this Xerox system*. Still 1 out of 100 wrong is far too many to send an automatic summons.

But maybe enough to send a warning letter or kick-out the image to a human operator.

*Yes, it would be ironic if a propped-up Xeroxed passenger face would defeat the algorithm...

SH
Slava H
Jul 24, 2017

The Xerox tests have been done by Xerox and was not confirmed in real life scenario by any 3rd party.

Its very hard to believe it, if to consider that no one else to best of my knowledge demonstrated with independent review such results.

Sandag test seems to be the only independent test in real live scenario http://actweb.org/wp-content/uploads/2016/04/CalTrans-Automated-Vehicle-Occupancy-Verification.pdf

New discussion

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

Newest discussions