Run OpenALPR on NVIDIA Jetson TX1

[Kobe Nein]

Is it possible to run OpenALPR real-time on NVIDIA Jetson TX1??

[Matt]

I ran it on a TK1 a year ago, and it was decently fast.  I bet a TX1 would be even better.

[Kobe Nein]

Thanks for your reply!

Matt於 2016年7月12日星期二 UTC+8下午9時13分13秒寫道：

[Vladimir Radnovic]

Kobe do u have TX1 ? do u test it ?

I am interesting too for it... to se perfomance now many FPS we can get .... whit CUDA..... I test whit OpenCL on SAMSUNG ARM* 8core...

tnx

vladimir

[Kobe Nein]

Not yet. I company will buy it recently

Vladimir Radnovic於 2016年7月14日星期四 UTC+8上午1時20分51秒寫道：

[Matt]

I compiled a special release of the OpenALPR commercial version for the Jetson TX-1.  It targets armhf and links against CUDA 7.0 using libopencv4tegra.  Since these binaries are able to drive the Maxwell GPU for plate detection, it is significantly faster.

You can pull these using the following commands:

wget -O - http://deb.openalpr.com/openalpr.gpg.key | sudo apt-key add -
echo "deb http://deb.openalpr.com/tx1-commercial/ trusty main" | sudo tee /etc/apt/sources.list.d/openalpr.list

sudo apt-get update && sudo apt-get -y install openalpr openalpr-daemon

# Enable the GPU detector

sudo su root

echo "detector = lbpgpu" >> /etc/openalpr/openalpr.conf

# Use all 4 CPUs for daemon analysis

echo "analysis_threads = 4" >> /etc/openalpr/alprd.conf

You will notice that this is slower when analyzing single images, but significantly faster when analyzing video streams, video files, or directories of images (due to GPU pipelining).  If you run the tegrastats utility while the OpenALPR process is running you should see the GPU usage (GR3D) close to 100%.

[Bara Muse]

Hi !

I'm evaluating some SBC solutions dedicated for OpenALPR, so trying to get the best price/openalpr-performance ratio.

It is my understanding that the only way to get fast LPR with those kind of platforms, is to get GPU-accelerated processing and as OpenALPR relies on OpenCV and OpenCV only have GPU support for CUDA, then only the NVDIA SBC platforms are eligible.

Now the question is : how faster is it ?

Compared to CPU-only platforms, like a rPI-3, way cheaper, and able to achieve ~15fps on a MJPEG stream (without too much optimisations, could maybe do better).

Power consumption is not really an issue, neither the volume.

What's important is the final price per camera stream.

Thanks a lot for your insights !

[Matt]

For some commercial customers I've tested on the TX-1 as well as the Odroid XU-4.  The TX-1 is, of course, faster.  But The ODroid is surprisingly fast for how inexpensive it is, since it has 8 ARM cores and supports OpenCL.  The Intel NUCs are also quite good, since the Intel consumer-grade CPUs have a GPU on-board that can be driven via OpenCL.  It's not especially powerful, but it helps like having an extra core or two.

Raspberry Pi 3 is too slow for anything faster than 5-10mph vehicles.

-Matt

[Marcos Luciano]

I have OpenALPR in my Rapsberry Pi 3 B but the performance is too slow. Is Odroid XU4 more powerful? How many percent aprox.? The investment is worth?

[Dave]

Any updates?

[Marcos Luciano]

Jetson Nano is slightly faster than Odroid N2 (OpenALPR won'tt use the CUDA GPU consistently, the GPU usage ranges 0%-100%-0%-100%...).

Odroid N2 is more faster than Odroid XU4.

Odroid XU4 is much faster than Pi 3B.

I bought all SBCs and tested it.

Em segunda-feira, 13 de janeiro de 2020 08:34:18 UTC-3, Dave escreveu:

Any updates?

[Marcos Luciano]

Jetson Nano = CUDA GPU

Odroid N2 = OpenCL

Odroid XU4 = CPU (OpenCL won't worked in my tests)

PI B3 = CPU (I think don't have OpenCL support)

[Dave]

The inconsistent GPU utilization is strange since they have a page dedicated to GPU acceleration here: http://www.openalpr.com/nvidia.html

[Marcos Luciano]

I think it's due to the tesseract OCR. Tesseract only uses CPU. OpenALPR Commercial cannot run on Jetson Nano, i've tested on free package. I tried few optimization options but GPU usage never stayed consistenty.

[Dave]

So the nano is the fastest sbc but it's inefficient? You've definitely followed these steps right? http://doc.openalpr.com/watchman.html?highlight=nvidia

[Marcos Luciano]

Nano have a high frame drop when plates is detected (Odroid N2 have similar frame drop). I'm using, today, the N2 to recognize plates in 3 camera streams, but sometimes recognize fails due fps drop (it's not 100% accurate).

Nano have a lot of dependecies erros when "install_nvidia" is selected on 'bash install' command. If you install agent without this flag, OpenALPR only runs with CPU. Changing detector to cuda, OpenALPR crashes. With opensource, all compile without errors.

[Dave]

Seems like the last commits by matthill were on Sep 19, 2018. I don't think we're going to get much help with this one :/

[Dave]

Email regarding the Jetson Nano FPS:

-------------------------------------------------------

Hi X,

Yes, The watchman agent does support the Jetson Nano up to about 10-15 FPS.

Best,

John

On Wed, Jan 15, 2020 at 5:12 PM X wrote:

Hello. I am looking to purchase a commercial license for the Watchman Agent software at $49 per camera/month.

I want to be sure that the Nvidia Jetson Nano  single board computer will be able to fully utilize the GPU acceleration capabilities OpenALPR has, as mentioned here https://www.openalpr.com/nvidia.html. Could you please confirm this?

The github for OpenALPR seems inactive. I'm not sure whether that's just the free version or commercial.

Regards,

[Marcos Luciano]