Analytics vs Rain: Axis, Bosch & Sony

BW
Ben Wood
•Published Oct 21, 2013 04:00 AM

****. **** *** * *** ********* to ********* ********* ****: **** **. ***** vs. ****. ******* ********** ***** ** ** one ***** *** ** ********* ** do **** ******* *** ******** ** night *** **** ***** ** *** rain?

** **** ******, ** ***** *** findings ** ******* *** **** ****** analytic ** ******** ** ***** ****. Here's **** *** *** ****** ****:

** *** ************/ ********* **** **' ** ***' away **** *** *******.

************, ** ****** * ********* ************* ** *** ****** **** **' to ***' ****.

**** *** *** *** ******** **** this ****:

  • ** ***** ********* **** ********* *** to ****** **** (********-**-*****)** ******** ******* **** ****** ********.
  • ***** **** *********** *** ******** ******* at **** ****** (** *** **** ~150') ** ****, **** ******** ** the ******* **** *** ****** ***** ********* ********. 
  • **** * ***** *******, *** ***** NBN-733V's *** *** **** ********* ********* ******* to **** ****** *********** ** ***' and ********** ****** *********** ** ***'. With ** **** **********, *** **** alert **** *** ***% ***** ***' *** 200' for **** ********* *** ***** ***, ************. Detection ******* ***** ** **** ***-*****'* VMF ********* ** ** *** **** at ***** ****** (**% ** **', 40% ** **' *** *% ** 100') ********** **** *** **-**** *******. 
  • **** *** ****** *** * *******, no ***** *********** ** *** ********* performance ** **** ****** **** ********, compared ** *** **** ** **-**** condition.  

***************

** ******** ** *** ******* *************** of *** ******** ******, ***** ***** show **** **** *** ***** ******* conditions **** **** ** ** ********* factor ** ******** **** ********* * reliable *****-**** ********* ******. *** ******* detection ******** **** ** ******* ** rain*** ** ****** ** ***** **** account, ******** **** ***** *******, **** as ******** ********, *** *** ******** conditions. ********** ********* ** ****** **** the ****** ****** ** ******** ** imaginary ******** ** **** ********* **** 100-150' ****.

*******, ***** ******** / ******, ***** understandably ******, *** *** *********** ********* problems *** ** *** ****.

****

** ****** *** ***** **** ********* in *** ****** ******* *********:

********

*****, ** ***** **** ******* * virtual ******** ***** *** ******* ***/********, as ***** ** *** ******** **** below. 

*** ******* ******* *** ******** ** repetitive ***** ** ******* ********* **** 20' ** ***' **** **** *** camera.

** **', **** **** ***** *** Bosch ***-**** ******** **** **** ********, while **** ***-*****'* ******* **** *** 7/10.

** ********* ******* **' *** **', success **** ** ***** ********* ************ ** ****** *** *********** ** 100'. ***** *** ***-**** ********* ** perform ******** ** **** *****.

** ***', **** *** ***** *** ***-**** started ** **** **** ***********. **** is ***** *********** **** **** **** compared ** ***** ******** ** **** range ** *** ******** ** **** test.

** ***', **** ** *** ******* triggered ** ***. ** ***** **********, the ***-**** *** ***** ** ** 100% ******** ***'.

********* ******* ***** ** **** ****** from *** ******* ******** ***** **** been ********* *****. ***** ***** **** 100% ******* ****, ******** ******* **** ******* ** ******. Since ** **** *** ******** *** false ****** ** **** ****, ******* rate ** ***** ** **** ****** in *** ******* *****.

Vehicle ** ********

** **** ****** *** ******** ** the********* ********* ** ***** *******, ********* ******* detection ** ***** *********; **', ***' and ***'. **** ***** *** ******* to *** ******** ****, *** ***** were *** ** ** *** ********, across *** ***.

******* *** **** ***** ** ***** NBN-733V ****** *** ** *** ******* activations ** *** ** *** ***** distances. *** **** *****, ** *** other ****, *** ****** *********** **** at *** ******* **** ***** ** 25', *** ********* **** ** ****** distances. ***** ******* **** ******* ** what ** *** ** *** ** rain ****. *********, ** **** *** observed *** ****** ** *** **** in *** ********* ***** ** ****** objects, **** ** ********, ***** ***'.


** ~***' ****'* ******** ***** ** ~50%.

*******, ** ***', *** ***-***** ** unable ** ****** *** ******* ** all.


*** ***** ***** ********** ******* *********** of ******** ** *** ****:

 

Comments (7)
Avatar
Robert Baxter
•Oct 21, 2013

Figen:

Thanks for the report. Over how many hours was the data collected? I have found that over extended periods of rain >12 hours (It rains a lot in Vancouver) you should get some false positives. When calculating detecting true events accuracy, how many true events were done per test?

Also, all of the testing was using crossing line detection, which I never use. I have a belief about crossing line detection as not being as good as "appearing in ROI". Early on I was finding that trespass by people would be missed because often fencing or hoarding would have storage containers/ obstructions beside them and entry might be obscured and crossing line detection would miss them, while appearing in ROI would detect them. Any thoughts on this?

Avatar
Ethan Ace
•Oct 22, 2013
In reply to Robert Baxter

Hi Robert, we tested over about 5 hours. Results were consistent all the way through, so we felt it was long enough to see patterns. However, we do believe there could be performance differences depending on the type of rain, depending on how heavy it is, how windy, potential for hail, etc.

We typically ran 10 attempts with a human subject at each distance. Vehicles, not as many, because performance was pretty close to predictable based on the first 2-3, due to the much larger object size.

This testing was limited to cross line detection for two reasons. First and foremost, because it is the most common analytic in use in the industry. Second, this is a continuation of our previous camera analytics test, which was limited to crossline.

We will be testing other analytics in the future (starting today actually), and will be using other rules in various conditions.

MB
Marc Beningson
•Oct 22, 2013

I did a major analytics deployment some years ago, and in the rain we had signficant issues with directional detection in the rain, especially in evening and dark. Headlight reflections on wet pavement on opposite side of the road on a curve made the analytics think there was movement in the wrong direction consistently. On an interstate there were hundreds of alarms each morning.

Doing tripline for our first effort was a good idea, but I'll bet other functions might be more adversely affected by rain and other weather conditions. I'm sure you will do follow-on studies of more advanced features.

Too bad you didn't test that other ananlytic that would have "learned" that rain is normal...

Avatar
Ethan Ace
•Oct 22, 2013
In reply to Marc Beningson

Hi Marc, we tested virtual tripwires as a continuation of our Camera Analytics Tested: Axis vs. Bosch vs. Sony report. We will be testing AgentVi, VideoIQ, and Bosch IVA in upcoming reports, so we'll be looking at other rules in varying conditions then. It is, unfortunately, very difficult to plan rain testing in Pennsylvania, as we've had one day of steady rain in the past month. But we're getting there!

MB
Marc Beningson
•Oct 22, 2013
In reply to Ethan Ace

What, you don't just look at your iPhone or Android app for weather until it says "rain?"

Sounds good. I would be very interested to see performance comparisons of integrated vs. add-on analytics, edge vs. server based, and the like. I've always thought that edge should be better, but of course the algorithms are not identical either.

I had some good success with AgentVI back when they were still Aspectus. I hope you will also include Sightlogix and Puretech, both of whom include geolocation which sometimes is a good adjunct and sometimes a good alternate for radar.

EV
Ed Vergara
•Oct 22, 2013

My initial thought was that the rain droplets on the lens would be the issue but according to this test the reduction in analytics performance looked more like the rain itself caused visibilty to be reduced which then reduced analytics performance. This tells me that fog would have the same effect.

Which brings up another question ... what is the effect on the analytics with snow? Would slow falling snow flakes near the camera be detected?

Avatar
Ethan Ace
•Oct 22, 2013
In reply to Ed Vergara

I would think the snow flake question depends on the size and speed of the flake. We'll be testing this when winter comes, assuming we actually get snow this year.

I believe that snow will have more of an obscuration effect than rain, similar to fog, which will have more of an impact on performance that the chance of false alarms from falling snow.