Strategy for LPR detection on video stream

358 views
Skip to first unread message

Nguyen Nam

unread,
Feb 8, 2021, 3:56:37 AM2/8/21
to doubango-ai

Hello group,
I am testing with UltimateALPR SDK from Github, it's fast. Used the default configuration, and there was some problem with accuracy. Also the github code recognize less plates than the web demo at doubango. May be I need more tuning.

My question is when I start with video stream instead of single image, one plate (vehicle) can appear in multiple frames during the time the vehicle passing over camera, which is expected. That lead to duplicate plate detection. Most of recognition was accurate, but few was wrong. What you think is best strategy to aggregate all those recognized plates of same vehicle, and then pick the rightest one for vehicle identification?   

I think someone had faced that duplicate detection before, and may come out with some advises on how to eliminate it?
Couple images was attached, taken 1 sec aside. Duplicate results.
frame-50.jpgframe-75.jpg
Regards,
Nguyen

Mamadou DIOP

unread,
Feb 8, 2021, 4:27:12 AM2/8/21
to Nguyen Nam, doubango-ai

Hi,

The cloud uses different models than the SDK. You can check the issue tracker (https://github.com/DoubangoTelecom/ultimateALPR-SDK/issues), the new models will be part of v3.3 to align the cloud with the SDK.

The main difference between the cloud and the default config in the SDK is the sensitivity for the pyramidal search. Set the value to 1.0 to improve the recall. More info at https://www.doubango.org/SDKs/anpr/docs/Improving_the_accuracy.html#golden-number

For the duplicate plates, use KCF tracker on the cars. This kind of tracker is implemented in OpenCV. We'll add it to CompV but car tracking will not be part of UltimateALPR but in UltimateITS only. A tracker will give you a unique identifier for each car, just save all detected plates for a car and keep the one with highest score.

Regards,

--
You received this message because you are subscribed to the Google Groups "doubango-ai" group.
To unsubscribe from this group and stop receiving emails from it, send an email to doubango-ai...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/doubango-ai/6476467f-f7c5-4cb6-b453-0d5bfb32bbefn%40googlegroups.com.

Nguyen Nam

unread,
Feb 8, 2021, 6:28:38 AM2/8/21
to Mamadou DIOP, doubango-ai
Thanks for your explanation. For UltimateITS, I guess I should contact your company for a demo SDK? Really want to check that tracker out. 
Regards, 
Nguyen 

Mamadou DIOP

unread,
Feb 8, 2021, 9:25:50 PM2/8/21
to Nguyen Nam, doubango-ai


On 2/8/2021 12:28 PM, Nguyen Nam wrote:
Thanks for your explanation. For UltimateITS, I guess I should contact your company for a demo SDK? Really want to check that tracker out.
For now UltimateITS is not avail for beta testing. UltimateITS is a very large project compared UltimateALPR and the target is cities with high budget.

Mamadou DIOP

unread,
Feb 9, 2021, 3:17:44 AM2/9/21
to Nguyen Nam, doubango-ai


On 2/9/2021 9:10 AM, Nguyen Nam wrote:
Thanks for  your reply. Going through all those object classification and tracking with opencv is overkill for this project. 

It's a pity that those features not available in UltimateALPR. Hope to see it   in the future.

The price for UltimateALPR is already very low compared to the competition. For example, the price for Android/RPI/Jetson is €39 per device one-time payment. For the same price you'll just have 1 month subscription for 1 camera if you're using OpenALPR or others.

We already have LPR, VMMR, LPCI, VCR, VBSR... features and adding a tracker to it without increasing the price will not make sense. I don't know if the competition have a tracker to merge the plates.

Regards, 
Nguyen 

Nguyen Nam

unread,
Feb 9, 2021, 3:45:34 AM2/9/21
to Mamadou DIOP, doubango-ai
Agreed on the prices :) Very reasonable indeed. And we don't mind to pay more for added features like object tracking. Please consider this. 

We are integrators, not CV/ML engineers, and really need something simple to use to get the job done. 

I don't know many SDKs, but another one we are testing from DTK (https://www.dtksoft.com/) has all things sorted out for developers. Just point it to the rtsp url, video files.. And they will extract plates, bounding boxes.. Guess they have object detection and tracking too, cause no duplicate in results. Seems slow tho. 

Regards, 
Nguyen


Mamadou DIOP

unread,
Feb 9, 2021, 4:16:18 AM2/9/21
to Nguyen Nam, doubango-ai


On 2/9/2021 9:45 AM, Nguyen Nam wrote:
Agreed on the prices :) Very reasonable indeed. And we don't mind to pay more for added features like object tracking. Please consider this. 

We are integrators, not CV/ML engineers, and really need something simple to use to get the job done. 

I don't know many SDKs, but another one we are testing from DTK (https://www.dtksoft.com/) has all things sorted out for developers. Just point it to the rtsp url, video files.. And they will extract plates, bounding boxes.. Guess they have object detection and tracking too, cause no duplicate in results. Seems slow tho.

From https://www.dtksoft.com/lprsolution I see they only have LPR. No VMMR, VCR, VBSR... Also they have a different target. We're more focused on developers.

I don't think they are using trackers on the car. Such solution is not easy. I have recommended trackers because this is the most robust solution but there are other possibilities:

    - For example, you can use the hamming distance on the plate number.

    - or if the car is visible, merge base on the vehicle info (make, model, color, body style...)

    - off course if you only keep plates with high scores (e.g. > 85%) you'll be almost sure the number is correct, you just need to use a dictionary with the number as key

Mamadou DIOP

unread,
Feb 9, 2021, 4:32:55 AM2/9/21
to Nguyen Nam, doubango-ai

According to https://www.zoominfo.com/c/dtk-software-company/346746356 their ANPR solution is 14 years old. UltimateAPLR is 1 year old (https://github.com/DoubangoTelecom/ultimateALPR-SDK/commits/master).

You can also check https://www.sighthound.com, they have a "ready to use" alpr solution.

Out next releases will focus on accuracy, then we'll add a dashboard.

Reply all
Reply to author
Forward
0 new messages