How much resolution is really needed for LPR

[wittaj]

We have folks come here all the time thinking that more MP is better, whether it be for general purposes or for LPR.

Those of us that have been around long enough know that sensor size is more important than MP.

Those that have been here awhile know that I share a representative sample of plates I get at night of vehicles traveling about 45MPH at 175 feet from my Z12E that is on the 2nd story soffit, My angle is about 40 degrees vertical, 50 degrees horizontal. Camera is 35 feet above street at this location.



So I was playing tonight and wanted to see how low resolution could you go before a 3rd party plate reader couldn't read it.

I am shocked, but this is D1 resolution substream from the Z12E at 175 feet away and the 3rd party plate reader has read every plate the same as the 2MP stream. The 2MP is a little more defined, but with the idea of getting plates is to bounce IR off a reflective plate, the differences are not as great as we may think.




Now I can see why OpenALPR says in many instance 720P is more than enough to read plates and saves a ton of CPU processing compared to 2MP or 4MP.

 

[wittaj]

So I thought I would show a representative daytime example between mainstream 2MP and substream D1 as it relates to plate reading.

Here is the 2MP version:



And the D1 version:



I have been running the 3rd party plate reader for 3 days on the lower resolution and it has read all the plates correctly.

So for the NOOB chasing MP, this help proves that sensor size is more important than resolution.

Now I am sure if this were really a D1 resolution camera, the results would not be the same. Unfortunately I do not have a D1 resolution camera with the zoom of the Z12E to compare it to LOL. Further, 2MP is my lowest resolution IP camera.

I don't know if an IP camera versus an analog camera would be a fair fight. I guess I can try a fixed lens 2.8mm 2MP fixed lens set to D1 and compare it to a Swann analog at D1 lol. In bright daylight it may stand a chance, but no way at night with an analog not being able to manually set parameters.

Another thing to add is that this is a D1 at 2046 bitrate, which at that resolution has a negligible impact to CPU% for those using BI. But it is interesting to see how high you can crank a D1 resolution and get fairly close to looking like a higher resource intensive MP image.

Further shows that since the 4MP 5442-Z4E costs the same as the 2MP 5241-Z12E, that one should favor the Z12E due to it having twice the optical zoom of the Z4E for LPR purposes.

 

[tech101]

How much Are you saving in CPU or say GPU resources on OPENALPR end with D1 vS Main ?

 

[wittaj]

How much Are you saving in CPU or say GPU resources on OPENALPR end with D1 vS Main ?

As you know, a GPU and OpenALPR was a joke. It was way more CPU% using the GPU than not using it.

With some other motion going on and some active plates, my CPU would get some peaks at 90% with mainstream. With D1 it doesn't go above 20%

 

[tech101]

That is a huge Diffrence from 90% to 20% damn I am missing out I guess need to dial in my zoom and focus to try to use D1 at-least for 1 LPR... Sounds like if not both..

 

[wittaj]

That is a huge Diffrence from 90% to 20% damn I am missing out I guess need to dial in my zoom and focus to try to use D1 at-least for 1 LPR... Sounds like if not both..

Yeah I was shocked that it could read it and do so well.

I cam across it by mistake LOL. I had my cam set to to use substream at 2MP at a lower FPS to try to lower the CPU.

For some reason when the camera rebooted it changed the substream to D1 and 512bitrate and it went a week before I needed a reason to look at OpenALPR and the aspect ratio was different and it didn't look as clear as I was used to, but it nailed every plate. So I figured hey lets try a higher bitrate and at that resolution the impact to CPU was not noticeable but the improvement to the pic was.

 

[tech101]

Thank you Wittaj its raining here but I did put my D1 To higher bitrate like you suggested for now, and left the zoom and focus alone.. Will see how this one LPR does. I was using D1 for BI anyways but at much lower bitrate but not at openalpr side...

 

[tech101]

Definitely D1 Is pretty darn amazing once I set Bitrate High.. Like you suggested For D1. I think I should pull the cam a bit up and zoom in even more tight .. Specially for those UPS truck who have there back plate like 3 or 4 feet high ... By middle of the Tail Gate door


View attachment TE-125.20230105_162201_1.mp4

 

[wittaj]

Definitely D1 Is pretty darn amazing once I set Bitrate High.. Like you suggested For D1. I think I should pull the cam a bit up and zoom in even more tight .. Specially for those UPS truck who have there back plate like 3 or 4 feet high ... By middle of the Tail Gate door


View attachment 150109

WOW - that is even more impressive than mine LOL.

Makes chasing MP look kinda silly now LOL.

 

[carteriii]

I love this analysis and optimization for those who can benefit from it.

Having said that, since we definitely have people new to LPR who aren't so technical and read these posts, I want to point out that if you are trying to capture license plates from 2 (or more!) lanes of traffic (as people commonly do), zooming out to have enough width for 2 lanes effectively requires 4x the camera resolution to get the same number of pixels for a single plate like the samples above. I hear that's why someone in marketing decided to call cameras "4k" rather than "2x". Marketing is probably also to blame for 720 and 1080 both being labeled "HD".

 

[TheWaterbug]

since we definitely have people new to LPR who aren't so technical and read these posts

That's me!

I'm quite technical in some areas, but not this particular one.

Does anyone make a dual-use camera that supplies a "normal" stream for regular surveillance plus an "LPR" stream with a much faster shutter speed and a cropped area of interest? Almost like a 2nd sub stream.

Given the very short frame time for LPR, a camera could either take discrete 1/1000 sec exposures alternating with regular exposures, or there are some modern sensors that can do "non destructive readout," wherein you can read the sensor multiple times without resetting the exposure. A lot modern sensors with programmable areas of interest, too, so the camera electronics could read out the "LPR frames" more rapidly than reading out the regular frames, which would help prevent the exposure/readout times from trampling on each other. And if the AOI is programmable, a smart application could even tell the sensor to track the plate as it moves through the overall FOV, to increase accuracy.

Or has this all been done already, but in cameras that I'm not seeing on amazon?

 

[wittaj]

The problem is the processors in these cameras do not have a lot of excess capacity and we have seen in cameras like the dual lens boobie cam that the little processor can't keep up, or if someone runs a camera at every rated spec.

So I think trying to get a camera that could quickly adjust fast and slow shutter to accomplish both LPR and overview visual would be difficult.

It would be better to dual lens the camera to do so and there are several out that under the right scenario could work (cars are closer to camera).

You could take the boobie camera and change out one of the the 2.8mm lens for a 25mm lens and could set the camera up that way if the vehicle was within 40 feet or so and then the other 2.8 is the oveview.

Other than that, I think the cost to put together a camera like that would be more than just buying two cameras and placing them side-by-side.

 

[TheWaterbug]

The problem is the processors in these cameras do not have a lot of excess capacity and we have seen in cameras like the dual lens boobie cam that the little processor can't keep up, or if someone runs a camera at every rated spec.

So I think trying to get a camera that could quickly adjust fast and slow shutter to accomplish both LPR and overview visual would be difficult.

It would be better to dual lens the camera to do so and there are several out that under the right scenario could work (cars are closer to camera).

You could take the boobie camera and change out one of the the 2.8mm lens for a 25mm lens and could set the camera up that way if the vehicle was within 40 feet or so and then the other 2.8 is the oveview.

Other than that, I think the cost to put together a camera like that would be more than just buying two cameras and placing them side-by-side.

If this hasn't been done yet, then this might be fertile ground for product development.

In my industry (mil/aero infrared imaging) some infrared sensors can dramatically change parameters from frame to frame, e.g. for two-color imaging (LWIR alternating with MWIR). I wonder if any consumer-grade CMOS sensors would be capable of changing gain/sensitivity on a frame-by-frame basis.

 

[wittaj]

I would hope the mil/aero infrared imaging uses better sensors and processors than these cameras LOL.

This camara is kinda close to that in that but it uses two lens - one with faster shutter in B/W and the other slower shutter to get color and then stitches it together:

Worlds First Review - Dahua DH-IPC-HFW5449T1-ASE+D2 - Dual Lens Full Color 2.0

Intro I’m pleased to announce we have another worlds first review to kick off, this time on the new HFW5449T1. So whats better than 1 camera (lens), well 2 of course. This camera makes some bold claims of Full Color regardless of low light environments utilizing 2 lenses. Lets jump into it but...

ipcamtalk.com

 

[richardgohth]

We have folks come here all the time thinking that more MP is better, whether it be for general purposes or for LPR.

Those of us that have been around long enough know that sensor size is more important than MP.

Those that have been here awhile know that I share a representative sample of plates I get at night of vehicles traveling about 45MPH at 175 feet from my Z12E that is on the 2nd story soffit, My angle is about 40 degrees vertical, 50 degrees horizontal. Camera is 35 feet above street at this location.

View attachment 119342

So I was playing tonight and wanted to see how low resolution could you go before a 3rd party plate reader couldn't read it.

I am shocked, but this is D1 resolution substream from the Z12E at 175 feet away and the 3rd party plate reader has read every plate the same as the 2MP stream. The 2MP is a little more defined, but with the idea of getting plates is to bounce IR off a reflective plate, the differences are not as great as we may think.


View attachment 119340

Now I can see why OpenALPR says in many instance 720P is more than enough to read plates and saves a ton of CPU processing compared to 2MP or 4MP.

300 pixels per metre

 

[TheWaterbug]

300 pixels per metre

LOL! This reminds me of way back in 2000, when I started managing a PR project with some product photography. These spiffy new digital cameras were all the rage, so I told the ads manager that our new photographer would be taking digital photos instead of scanning film prints. She warned me several times to be sure the new photos would be "at least 300 dpi," so I wrote her an extensive email explaining how "DPI" was no longer a meaningful metric with a digital camera and fully digital workflow.

We went back and forth for many, many emails, until I finally had some confidence that I had explained it to her. She finally said, "Ok. . . . . as long as they're at least 300 DPI."

And now we're back to 300 ppm.

 

[richardgohth]

LOL! This reminds me of way back in 2000, when I started managing a PR project with some product photography. These spiffy new digital cameras were all the rage, so I told the ads manager that our new photographer would be taking digital photos instead of scanning film prints. She warned me several times to be sure the new photos would be "at least 300 dpi," so I wrote her an extensive email explaining how "DPI" was no longer a meaningful metric with a digital camera and fully digital workflow.

We went back and forth for many, many emails, until I finally had some confidence that I had explained it to her. She finally said, "Ok. . . . . as long as they're at least 300 DPI."

And now we're back to 300 ppm.

not relevant . whats your point?
The stroke width of car number characters are typically 1 cm.
Minimum nyquist criterion wld be 2 pixels per stroke width / cm to discern the stroke.
A resolution of 3 pixels per stroke width would be desirable to provide for some motion blur, out of focus and exposure issues.

Post reminds of when I started coding on the IBM4341 way back in 1980s

 

[TheWaterbug]

not relevant . whats your point?

Nothing really. The number "300 <insert units here>" as a recommended resolution just stuck out in my mind, since it used to be a key metric, then it became a meaningless metric, and now in this case it's an appropriate metric again.

 

[TheWaterbug]

I would hope the mil/aero infrared imaging uses better sensors and processors than these cameras LOL.

Yes, but a lot of features from the high-end sensors trickle down to commercial-grade and then consumer-grade sensors, quickly. For example, back-thinned sensors used to be super expensive exotic devices restricted to spy satellites and space telescopes. Now every iPhone has several of them.

 

[wittaj]

We are talking here about camera manufacturers that still use Internet Explorer as the browser they write firmware around LOL

When we have seen folks here try to run a camera at 30FPS and the highest bitrate possible and the camera is choking and only becomes stable when they drop the FPS and and bitrate, I think we have a ways to go before we see this alternating fast and slow shutter lens LOL.

Plus keep in mind us homeowners are not their target market and most of their customers that run true ANPR cameras probably have enough light to be in color, so they don't really have an incentive to try this concept as their customers can get great captures of the plate and the vehicle in color.