Any specific tips for successfully capturing Motorcycle License Plates?

[alekk]

[Note: Some Hikvision specific questions, but I thought the LPR forum might be better to post in]

Summary: I've been using a couple of LPR camera's to successfully capture license plates in our neighborhood for several months. As part of that, I did a fair amount of stress testing at our two gated entrances and believe the current setup is close to 100% successful in daylight and better than 90% at night. But I just realized that in the thousands of images I've look at, the LPR system hasn't captured any motorcycles which have smaller license plates. Any tips/suggestions on why that might be?

And specifically, are there any disadvantages to setting the Hikvision "Plate Mode" from Large to Small?

Details/History:
The camera's are HNC512-VBA32-LPR from Nelly Security running their most recent (4/2/2019!) firmware of 5.4.5. I'm pretty sure this camera is just a re-labeled Hikvision DS-2CD4A26FWD-IZHS8/P that has an 8-32mm lens with a (big!) 1/1.8" sensor with (wide/fast) F/1.4 lens. I.e. while there are newer models, the basic image specs are pretty solid ... and again, my results with cars are very, very good (although it sometimes will record an "O" as a "0", etc.)

While not very common, I happened to see a motorcycle in the neighborhood a few days ago ... and "DUH!" ... I've NEVER see a motorcycle in all of the captured license plate images I've looked at - D'OH! ;-)

So I went "looking" for that motorcycle - the NESTcam shows "action" at the Front Gate ... but the LPR software did NOT capture a plate ... although if you look at the raw video footage, it's easily readable with the Mark I Eyeball in multiple frames ... and seems to me that it should have captured that plate (?) - see attached image.

I got some suggestions (see below) from Nelly's (their support has been great) but I'm reluctant to implement them since the current setup works great for us.

So I'm specifically wondering about the "Plate Mode" setting which is currently set to Large. The other option is Small. The Hikvision documentation about this is very sparse (basically how many pixels on target - we have plenty) ... but doesn't specify what possible disadvantage there might be setting to Small would be. Nelly's wasn't able to add anything (nor can I find anything by Googling) so I suggested posting something at ipcamtalk and they said go-for-it ... and said "the guys over there are super knowledgeable" ... which is my experience as a long-time lurker.

Thanks,
alek

P.S. Some suggestions and why I'm reluctant to implement them - again, the success rate is very high with the current setup for cars, trucks, etc.
1. Tighter Zoom: I'm about as tight as I can get and still see at the top which way the car turns which I think is useful info. There are plenty of pixels on the plate.
2. Make License Plate Detection Area Smaller: I'm reluctant to do this since due to sun reflection, bike rack, and other factors, the license plate may not be readable in certain areas as the car passes through.
3. Set Frame Per Second to 60 instead of current 30: I was surprised by this recommendation since I pointed out that cars have to slow down for the gate ... and in my testing when I have gone SMOKING through the gate, there were still plenty of frames to capture an image. Setting to 60FPS would reduce our storage time ... and interestingly enough, the 60FPS setting is in an unusual location and is actually removed in a subsequent release of firmware for the OEM camera ... so I think there are possible issues with this.

 

[wittaj]

I use a non-ANPR camera and use OpenALPR and get motorcycles, so it is absolutely the settings that you are mentioning changing in an ANPR camera.

Tighter zoom works. If you care about the way the car turns, then add a cheap overview camera. The whole point of an LPR camera is to get the plates, not tell which way a car turns. I would rather get more plates.

You could make the plate mode setting small. I suspect that the AI of the camera sees this small plate and it doesn't meet the minimum size criteria for the large setting - so it skips it.

As you have seen, increasing FPS isn't the answer. Cars pass mine at 45 MPH and I am running 8 FPS. Reduces storage, less CPU for OpenALPR, and it captures all the plates. Shutter speed is much more important than FPS.

 

[Flintstone61]

impressive. 45 mph @ 8FPS. My noob parking lot setting are for 15-20mph cars @ 12 FPS. I think my shutter is at 1000. I'm afraid to fool with it because it's working so good.

 

[alekk]

@wittaj : I hear 'ya on even tighter zoom since LPR is for LPR and that we COULD have another camera to see which direction the car turns. But with the current setup, there are plenty of pixels on target (see image below) and it's nice to get a "two'fer" in the same video. We are running at 30FPS ... and in my testing, I've blown through the gate at 30+MPH (using the remote to open the gate ... and assuming (!) it will open before I hit the gate!) and even at night, I get the plate most times. Nice to hear that even at 45 MPH and 8 FPS, you are getting plates.

@Flintstone61 : Ditto comments above and also that similarly, it's working well so I'm reluctant to change things. FYI that shutter speed is 1/400s.

So does anyone have any experience/knowledge of the pro/cons of setting "Plate Mode" to Small versus Large?

P.S. I agree @wittaj that perhaps the algorithm is saying the plate is too small for the criteria. The key question is why would I NOT set Plate Mode to Small?

 

[samplenhold]

The key question is why would I NOT set Plate Mode to Small?

I would imagine that the only real way to know is to try it. What is the worst that could happen? If once you set it to small you start missing car plates, set it back. If you are only getting a few MC's, you may have a long wait to find out if it works.

Other possible problems might be miss-tagging of plates or taxing the CPU looking for small plates?

 

[alekk]

I agree at the end of the day, I'll just have to try/test it. But I was hoping to understand what the pro/cons are of Large/Small.

I just posted in the Hikvision forum (since this is a vendor specific question) with pointer to this thread in case any of the experts over there can chime in.

 

[alekk]

So I switched Plate Mode from Large to Small ... and based on historical data, I don't think I'm missing any plates. In fact, the numbers are slightly higher, but I think due to slightly more duplicates - i.e. the same vehicle with different variations of plate alphanumerics - funny "DOGPOOP" example in this thread. Regardless, as best I could tell, there does not appear to be an disadvantage to setting Plate Mode to Small.

However, flipping through a couple of thousand images, I never saw a motorcycle ... but then again, they are really rare in our neighborhood. So I need to find someone with a motorcycle who would be willing to do some testing for me!

There does seem to be adequate resolution. The first image is a (snazzy!) car with the approximate height of the "plate letter" being 30 pixels. The 2nd image is a riding lawn mower (!) that the LPR software captured ... and it got the numbers right even though they are only 12 pixel high. The 3rd images is that motorcycle where the letters are ~17 pixels high ... but this was BEFORE I switched "Plate Mode" to Small. So again, I need to do some real world testing with a motorcycle.

BTW, the HELP info says:
SMALL: height of a license plate should be 20 to 30 pixels in the image captured by a 2 megapixel resolution camera.
LARGE: It means the height of a license plate should be 30 to 40 pixels in the image captured by a 2 megapixel resolution camera.
So probably good I made this change.

 

[alekk]

Summary update is that setting Plate Mode to SMALL did not appear to make any difference ... but I could not "see" any downside, so I'm going to leave it as-is.

A reminder that motorcycles are very rare in our neighborhood, but I happened to see a motorcycle drive by my house the afternoon of July 4th ... so I checked the footage ... and SCORE(!) ... the LPR Software correctly captured the plate as seen in the first image attached and here's the matching video. However, that was the ONLY one captured over a two week period of over 5,000 images.

There actually was one other motorcycle "captured" by the LPR camera ... but this is because it took a snapshot of the following car as seen in the second image - third image is from the matching video - plenty of pixels to get a good plate, but maybe the algorithm got confused by the waving flag.

So until I've got someone willing to "test" things with a motorcycle, I'm just going to accept this limitation.

 

[bigredfish]

Im not using any automated plate capture program, with 30 days storage on a small cul-de-sac street, if we have a problem I'll easily be able to find it given a day and time window.

But thought I'd post this from last night as its a good capture for night at 100+ft

From an older 5231Z-12E