Challenge Accepted - 4K License Plate Recognition @ 85Ft with 4K S3 Camera (EmpireTech B58IR-Z4E-S3)

[Wildcat_1]

All, hope everyone had a great weekend. I wanted to post this video I took using the new S3 cams (EmpireTech B58IR-Z4E-S3 4K Bullet Camera, Dahua HFW5842E-Z4E-S3) I’ve been testing. Specifically I was challenged via some forum members and YouTube viewers to push these new cameras (4K variant) to the limit and attempt to use them for LPR (NPR) recognition. To be clear this is not to be confused with Automated License / Number Plate Recognition which as I’ve shown before has to be supported by the camera through OCR processing as part of the AI algorithm. These cameras actually DO support that BUT critically this has not been implemented yet in FW (read my bug list in the main review of the 4K Turret version HERE).

I’ve shown before and most people are aware that 2MP, 4MP cams at Z4 or Z12 do a good job in these kinds of situations BUT as I’ve mentioned before, light requirements, noise etc increase dramatically for 8MP/4K, making this a harder task.

So with that said, in this video I want to show you what’s possible, how you can (and MUST) tame the over-sharpening and not only demonstrate the results but share the settings I put in place when undertaking this challenge, including making it more difficult on myself. Unfortunately YouTube adds compression artifacts but I've tried to ensure those are kept to a minimum as always.

In case you're asking, YES, these do great for LPR and even ALPR (using on NVR AI, for the moment), during the day.

Hope this helps, enjoy ! Please do remember to like & subscribe to the video, it does help at least get the content promoted in the algorithm and these videos do take a LOT of time to produce.

Thanks

Do remember to select 4K for playback.

 

[wittaj]

WOW Great video as always.

I guess I need to try mine out and see if I can reproduce your results LOL.

 

[Wildcat_1]

WOW Great video as always.

I guess I need to try mine out and see if I can reproduce your results LOL.

Hahah, thanks for the kind words as always. It really was an interesting one for me with these cams. As you all know, I dial systems in on a regular basis but these cams DO need a LOT of tweaking in the sharpness, gamma, areas which was the real part of this challenge. I should (could in future) do a straight out of the box default version of this video BUT to be 100% transparent as always, these are NOT good out of the box without tweaking, especially for LPR.

Have fun with yours and hopefully my settings will serve as a good starting point for you @wittaj.

I also intend to do a video showing AcuPick 2.0 in action on the XI NVR's so will let you all know when I get that ready. However, next focus is on the bug list and implementations for Dahua to fix the 5 areas I'm tracking. Want to get these cams as close to 100% for everyone, ASAP.

 

[wittaj]

Hahah, thanks for the kind words as always. It really was an interesting one for me with these cams. As you all know, I dial systems in on a regular basis but these cams DO need a LOT of tweaking in the sharpness, gamma, areas which was the real part of this challenge. I should (could in future) do a straight out of the box default version of this video BUT to be 100% transparent as always, these are NOT good out of the box without tweaking, especially for LPR.

Have fun with yours and hopefully my settings will serve as a good starting point for you @wittaj.

I also intend to do a video showing AcuPick 2.0 in action on the XI NVR's so will let you all know when I get that ready. However, next focus is on the bug list and implementations for Dahua to fix the 5 areas I'm tracking. Want to get these cams as close to 100% for everyone, ASAP.

Yes it does need a lot of tweaking. More than any other camera I have had. I ended up using it as overview because I could never find the right combination, but this gives me hope maybe I can repurpose it to an IDENTIFY camera with some tweaks starting at your numbers.

 

[Wildcat_1]

Yes it does need a lot of tweaking. More than any other camera I have had. I ended up using it as overview because I could never find the right combination, but this gives me hope maybe I can repurpose it to an IDENTIFY camera with some tweaks starting at your numbers.

Yes, my goal was to see how far I could push the Z4 variant and like you, initially I thought this was going to be an impossible challenge only due to the sharpening and artifacting / noise seen out of the box. However, I knew I could do something with it so set the task of dialing in over 3 nights. Hopefully that work, helps you and others. These cams really are great and with the bugs I'm working on, hopefully being taken on and resolved by Dahua, will make them a top pick for deployments in 2024.

 

[Parley]

Good review. As I have stated before, I am using Andy's IPC-B541R-Z4E-S3 out to at least 85 feet on one camera, maybe two. These are California plates, and they work just fine at that distance. These are 2K cameras and not the 4K ones that you used.

​

 

[bigredfish]

Nicely done sir

Can you get a snap of the area in regular color mode or even phone so we can get an idea of ambient light

 

[Wildcat_1]

Nicely done sir

Can you get a snap of the area in regular color mode or even phone so we can get an idea of ambient light

Sure, it won't be much of a pic due to the lack of light but will aim to get one over the next few days for sure

 

[bigredfish]

Was really just wondering if any close by street lights or houses with big lights near the capture point

 

[EMPIRETECANDY]

Thanks a lot for @Wildcat_1 great trying, People know The 4MP IPC-B54IR-Z4E S3 well, it's working great at low light place as a LPR, and no one use 4k, but wildcat make this happen, great job!
Here is wildcat's setting for the 4K.

 

[steve1225]

I have one BIG question (elephant in the room) to @Wildcat_1 and all other people doing LPR this way:

Why You do 1/2000 shutter at night capturing only lights and IR reflection from plate numbers?

Why You don't do much faster shutter (1/100 - 1/250 at night) with WDR >= 50??

FULL WDR (with values >= 50) on 5442/5842 is doing triple/double flash.
So for each frame camera is taking 3/2 shots, where first one in normal, second is much FASTER (something like 1/16 of first one - so if first is 1/100, second will be 1/1600), third one is even much faster..
Then camera joins all shots using WDR/HDR algorithm - most scene is taken from first long shot, but very bright elements (like light or plate reflection) is taken from second/third..

Car plate are very reflective to IR at night. So they overexpose image at slower shutter times (like 1/100) and camera configured for FULL WDR are taking car plate from second shot taken at much faster speed (like 1/1600)..

In this scenario You see car, plates, people, road - all elements.. And plates are sharp (because they are taken from much faster second/third shot)..
In this scenario You have also working FULL Video MetaData or IVS.
Cameras see cars/plates/peoples/faces - so it can detect / capture them and do snapshots / visual analysis (VMD attributes).

I think that WDR will solve problem of different plates in the USA - some are much brighter, some are very dim (IR reflection)..
In normal backlight mode this can be a problem, but not for WDR...

Example of night captures taken with 5442-Z4 at 1/100 with IR in WDR (70) mode...
AcuPick disabled, FULL Video MetaData selected as SmartPlan, no IVS, no SMD...
Plus AI Coding + ABR to remove some compression artifacts..
Most other camera image config values are DEFAULT on this camera..
Video MetaData received by 5232-EI (but any 5xxx NVR will work here - from 4KS2, -I, -I/L, -EI, -XI lines).

You see everything... And camera is doing not only vehicles but also people (AI Search->Human detection is also working)...
The only minus is small chroma aberration around plates (there is small time gap between first and second flash and car is moving)..

ps. this even works better with 7442-Z4 where You have Time-Division Exposure technology for VMD and different video algorithms... Plus working full ANPR (plate numbers are OCR'ed and full searchable on NVR)...

 

[duplo]

Why You do 1/2000 shutter at night capturing only lights and IR reflection from plate numbers?

Why You don't do much faster shutter (1/100 - 1/250 at night) with WDR >= 50??

1/2000s = 0,0005s
1/100s = 0,01s

slow road)

Maybe this is what you dont understand.

Slow road , what ever you mean with "slow" ...? 5, 10, 20, 50 ? mph, kph?

Maybe you should post a comparison to prove your statements at reasonable speeds like 30 kph


Also maybe depends on camera. Some cams ive used in the past tend to be like gain 100 with wdr over 30 at night

 

[steve1225]

1/2000s = 0,0005s
1/100s = 0,01s

Maybe this is what you dont understand.

Read what I wrote later in the post about how WDR works... by taking 3 shots per frame with different shutters time (second much faster, third even more).

Slow road , what ever you mean with "slow" ...? 5, 10, 20, 50 ? mph, kph?

Maybe you should post a comparison to prove your statements at reasonable speeds like 30 kph


Also maybe depends on camera. Some cams ive used in the past tend to be like gain 100 with wdr over 30 at night

this works also for faster roads (You start with lower shutter like 1/200 or 1/250)...

You chose wrong WDR value - there are two WDR modes, one LITE version - single flash with values <50 (sometimes <40 or <45 depending of camera model), second FULL version - with double/triple flash with values >= 50..

We are talking about FULL WDR - so with double/triple flash - configured with WDR values >= 50..
Only in this mode this magic works - full image taken at 1/100 (or 1/200) are joined with very bright elements taken at 1/800 or 1/1600 or 1/3200 depending of algorithm and camera model.....

 

[duplo]

Its hard to follow your line of thought

You generalizing and mixing many things together and still not say what model (s1/s3) / what speeds you are showing here.

Even a reolink camera can capture a license plate at night with automode on still image and <1kph/mph under 5m.

There are some models which fail with any WDR settings even at day, because the overlay is not perfect and will result in lost in sharpness, really bad colors etc.

Axis Lightfinder 2.0 use WDR techniques which work quite well. In my observations Dahua fails in the past.

 

[steve1225]

Its hard to follow your line of thought

You generalizing and mixing many things together and still not say what model at what speeds you are showing here.

5442 is doing triple flash (WDR 140db) - one of the unique features of this model...
5842 is doing double flash (WDR 120db) - as 95% cameras on the market..

if base shot is 1/100 (0.01) then second one is faster by 4-8-16x..
So it can be 1/400 (0.0025) or 1/800 (0.00125) or even 1/1600 (0.000625)..

How many - I don't known - this is camera model / algorithm depended..
Hidden by Dahua..

But exact shutter times are NOT important.
It is important here that the camera takes two or three shots for each frame in full WDR mode.
And each shot is taken with VERY DIFFERENT EXPOSURE TIME to capture elements with different levels of lighting.

We use fact that car plate are very reflective to IR at night. So they overexpose image at slower shutter times (like 1/100) and camera configured for FULL WDR are taking car plate from second shot taken at much faster speed (like 1/1600 - but we don't known what exact values is hidden in WDR algorithm for specified camera model)..

This don't work at day or without IR. And require FULL WDR (>= 50)...

Then you saying faster shutter and obviously mean slower by the numbers you are telling.

1/200 (0.005) is faster shutter time that 1/100 (0.01)..
1/1000 (0.001) is faster shutter time that 1/200 (0.005)..

Even a reolink camera can capture a license plate at night with automode on still image and <1kph/mph.

One more time:
THIS WORKS FOR CARS AT NIGHT AT ANY NORMAL SPEED!!!

And this works only with WDR>=50 and require IR (strong reflection)..
So this don't work at day* - then You need much faster shutters times..

Yes, sometimes depending of camera model, light situation and car speed You need to start with different value that 1/100 (0.01). So I wrote shutter time 1/100 (0.01) to 1/250 (0.004) as starting value for night configuration.

* - not always, especially on 7442

There are some models which fail with any WDR settings even at day, because the overlay is not perfect and will result in lost in sharpness, really bad colors etc.

Axis Lightfinder 2.0 use WDR techniques which work quite well. In my observations Dahua fails in the past.

Lite WDR (<40) or FULL WDR (>= 50)?

WDR is NOT for having BEAUTIFUL picture with BEAUTIFUL colors.
Full WDR is for taking picture where You have all elements VISIBLE and SHARP regardless of strong changes in lighting (very dark areas vs lights directed at camera).

Full WDR is for taking pictures which are EVIDENCE for police.
Axis is naming WDR as Forensic WDR.. Which are describing what Full WDR is.

 

[bigredfish]

NOT FAIR!
Euro Plates I can read those from Florida

 

[steve1225]

NOT FAIR!
Euro Plates I can read those from Florida

I've been many times in the USA and I known that Your car plate situation is "fucked up"
But if doing very strong IR + very short exposure time are working for You - it mostly catches plates..
Then doing very strong IR + two exposures - one long and very short will work the same - except it will catch everything else (car, humans, streets)..

FULL WDR algorithm will combine longer exposure with everything in frame with short one You are doing now - which contain only lights and reflected plate number..
In both cases we use strong IR reflection from car plates, which overexpose the photo.

One more thing:
I do FULL Video MetaData on both 5442-ZE and 7442-ZE...
In this mode both 5442 and 7442 have enabled FULL ANPR engine inside...
Engine which analyzes both car image and plate number and optimize some internal camera parameters to work better.
Engine which try many shots of car plate and choose one which is most sharp.

@Wildcat_1 wrote many times that 5442 have enabled ANPR but the result (OCR'ed car plate number) is simply not saved (market segmentation - ANPR is reserved for 7xxx line).
That's true - this engine (part of FULL Video MetaData) is somewhat critical to get best results from WDR on longer exposure times..

SO: Simply try..
In AI -> SmartPlan disable AcuPick (camera will reboot) and select Video MetaData.. Default Video MetaData parameters are OK...
In Camera -> Image settings do another image profile (not day/night, but general or front light) and configure shutter priority - manual - time 0 - 10ms (or 5ms or even 4ms)..
Enable WDR with values >= 50 (60-70 works usually best)..
Then redraw time switch settings for actual month - draw selected profile (General or Front Light) for night's hours (overdraw night profile in time switch settings).

Simply try - for night or two.. You can play with max exposure time (0.01 / 0.005 / 0.004)..
You can play with WDR values (everything >= 50)..
You can play with all Illumination settings (automatic mode and manual mode, in most cases 100% FULL IR POWER will be to MUCH)..

You can check results on 5xxx NVR - AI Search -> Motor Vehicle detection, You will get very nice snapshots of cars and plates..
If something will be wrong then You try to play with Camera->Image settings for a new profile...

If You will be not happy with this new mode/ settings - then You can come back in Camera -> Image to Old Night Profile (redraw back time switch settings)..

 

[Wildcat_1]

Couple of notes:

1. You have to be careful when you're mentioning 5442, remember that series has been around forever. ANPR is NOT supported, NOT implemented on the S1 or S2 models of 5x42 cameras, period. However the new S3 cameras DO support ANPR (I showed how you can check the cam capabilities and what is being tagged in earlier posts) BUT critically has NOT been implemented yet. This is a key difference between S1, S2 and S3 versions. Now whether Dahua listens to my feedback, my bug report and completes implementation remains to be seen. However these cams DO support it, 100%. As I mentioned before, IF Dahua decides not to fully implement in FW then that is of course their choice and as I also mentioned before this is generally not to cannibalize from other models such as 7xxx or dedicated LPR traffic cams like I've reviewed before.
   
   
2. What is mentioned above about 1/100, 1/250 and WDR is VERY situation and most important location and target dependent. YES, there are situations where if you have controlled choke points such as Entry/Access areas OR are lucky enough to control vehicles coming directly towards camera then in those situations you 100% can adjust down accordingly. I've shown this on MANY traffic camera reviews and videos. However, for general street deployments where you don't have ideal angles, you do have poor to no lighting, vehicles are traveling about 25 to 45 or higher miles per hour then you do have to deploy differently. Changing any one of these criteria 100% can assist. In other words, slower vehicles, more lighting (ideally distributed), better angle, lower height etc BUT unless you are deploying in those type of situations then using settings that increase your odds for the multitude of vehicles (fast, slow, hogging the middle of the road vs one side or other, large, small) and plates (high contrast, low contrast, black on white, white on color, dirty, clean) etc is where you benefit, as shown from the settings I've featured across this video. As I mentioned before thats the beauty of profiles on the V5 web UI is that you can setup your day/night and individual profiles which works really well when you have situations that benefit intense dialing in for target and location requirements that are different between times of day.
   Now, as I call out, in those cases, 100% you should be deploying (to be fair you should anyway as I’ve mentioned for years) secondary and tertiary cameras (fixed wide, PTZ, both). This is the best of all worlds and using PTZ activation through NVR or NVR middleware as I’ve demonstrated before, takes your system to the next level.
   
   
3. WDR (heck, any backlight mode) you have to be INCREDIBLY CAREFUL with using, especially in higher numbers. Why ? because it adds a lot of noise and creates motion blur . If you add this to the slower shutter speeds (1/100 or 1/120) most times you'll run into some issues. 1/250 which is the minimum I have and would ever recommend for vehicles (outside of course of controlled barrier entry points for example) is still susceptible to this based again on vehicle speed and angle of deployment. So use backlight modes, sparingly. My recommendation here is do NOT use them unless you have to.
   
   
4. Always keep in mind that LPR/NPR/ and especially ALPR/ANPR is VERY dependent on a number of location, lighting and target attributes. This is why we always state this is an art and no 1 situation or deployment is the same as another which is why we all deploy specific to those target and location needs. The S3's are powerhouses for their series and cost. I’ve successfully deployed these for both day and night, LPR and ALPR (using the NVR) and the caps are great. Seeing 100’s of these now and the database on those installs is looking good
   
   
5. IR strength again will come down to distance to target, IR illuminators in use, light on scene and angle to plate.
   
   
6. As I've mentioned, definitely you can benefit from AI on NVR for units that support it. For example the XI, heck even the older IL support this too and you can use those for ANPR caps with the cams too.
   
   
7. AcuPick 2.0 should not be dismissed however as it does feature (and will increase further) some very useful benefits such as Map Tracks is a very cool feature that I would like to see implemented in other areas of code. Tracking an identified target across 30 cameras with the click of button is something that should be part of regular operating features on the NVRs but at least it’s there in some form now which is great.
   
   
8. Remember there is a big difference between x842 and x442 cameras in terms of capabilities, functions and critically, light gathering as well as noise ratio.
   
   
9. As Steve and I have both said, we’re advocates and I have been for a long while on Video Metadata, that should be used most of the time. However, remember that as I mentioned toned before AcuPick 2.0 DOES also feature the Experience Database which is a great addition and super useful (will do a video on this too) which you lose IF you use Video Metadata only.

BTW Steve definitely update your settings on that cam in the pic above, leaving them all at 50 on these S3’s is too sharp and introduces noise as I mentioned and showed. Also, no one should be running 2D or 3D NR at 50 !

That leads me to one of the bugs I am filing too. When you switch off AcuPick mode it wipes out the image and exposure settings on reboot. This may be why your cam is all on defaults. Either way pointing that out as that is 1 of a number of nasty bugs I am currently tracking.

I’ll be doing more videos both on the fixed lens, the bug video + an NVR walkthrough on AcuPick 2.0 as well.

HTH

 

[Wildcat_1]

Was really just wondering if any close by street lights or houses with big lights near the capture point

Don't worry Cinders, you shall have the picture Seriously though, great request. I'll look to upload when I can

 

[samplenhold]

But most of us are needing longer focal lengths, like Z12E and not Z4E. I am running 45mm and 52mm on my two LPR cams.