Thanks Mike. I have changed the settings per your suggestion. I will let you know the results once I see some action on the license cam.
New CodeProject.AI License Plate Recognition (YOLO11) Module
- Thread starter MikeLud1
- Start date
When I tried these settings, nothing was captured. No responses were received from License Plate Recognition (YOLO11.NET) 1.1.1 or Object Detection (YOLO11.NET) 1.3.0. Absolutely no alert clips were recorded.
Vettester
Known around here
- Feb 5, 2017
- 1,024
- 1,004
Something is still wrong then. Instead of wasting time trying to troubleshoot the issue I found that it’s quicker just to remove everything and start over.
I ended up deleting everything and reinstalling codeproject and the older modules for now.
I hope no news is good news. To all how is License Plate Recognition (YOLO11 .NET) 1.1.1 module working for everyone?
Like gangbusters! If my camera can see it, ALPR reads it. Fantastic job Mike!
Mike, I reported last Friday that License Plate Recognition (YOLO11 .NET) 1.1.1 stopped working for me. I tried the Blue Iris settings you suggested, but it didn't help. As a result, I reverted back to License Plate Reader 3.3.4 and Object Detection (YOLOv5.NET)1.14.0.
Make sure when you remove the non-working modules that you delete their whole folder out of the modules folder and when installing the modules flip the switch "do not use download cache".
After you install fresh you can stop/start CPAI or restart the whole machine. After that detections should work fine.
Before I reverted back to the older versions. I did just as you describe. I can try it once again, but there is no reason to suggest that it would start working now.
It's a bit finicky when first starting up, check the server log for errors during startup they are pretty good at identifying any issues. If there are no errors then just give it 5-10 minutes to start detecting things.
This will be easier if you can post a snapshot of your log so we can get an idea of what it is doing. No detecions vs. nothing recognized vs. nothing at all.
After almost a week of use, here is what I can report:
The Good:
I've increased the confidence level in BI to reject the phantom reads, but would be great if the confidence level was a little more linear. Overall, this is the biggest improvement to ALPR accuracy for me, after make a lot of changes to camera, lighting, BI tuning, etc.
Would love to start getting vehicle identification data (color, make, model, type, etc). Speed would be nice, but less important for me.
Thanks again Mike!
The Good:
- Accuracy is much better than the old version, approx 35-40% less read errors
- Good performance on iGPU (Intel 11th gen) avg 77ms inference time
- No errors in the log! (this wasnt the case with previous version)
- Ignores symbols in plates (handicap, heart, logo, etc)
- Overly confident, a lot of misreads show 100% confidence
- Below 80% confidence = invents plates
- 80%-90% = 100% chance of a misread
- Common misreads
- 0 and O
- I and T
- 4 and L (way better than previous version)
- Plates with small vertical letters
- Plate detection is aggressive
- If it sees text that it will want to identify it as a plate
- Vehicles with lots of decals, sticker, etc tends to generate misreads
- DOT and CA numbers are read as plates
- Phone numbers on sides of trucks are read as plates
I've increased the confidence level in BI to reject the phantom reads, but would be great if the confidence level was a little more linear. Overall, this is the biggest improvement to ALPR accuracy for me, after make a lot of changes to camera, lighting, BI tuning, etc.
Would love to start getting vehicle identification data (color, make, model, type, etc). Speed would be nice, but less important for me.
Thanks again Mike!
Attachments
Most states and counties do not use the letter O. The dataset I used to train the OCR does not have the letter O so if there is a letter O it will always read as the number 0.
Link to the dataset used to train the OCR
