New built in AI LPR in Blue Iris
- Thread starter International784
- Start date
I see your question has been sitting unanswered for a while, but it's related to something I came here to post about today, so I'll try to answer it even though I suspect you may have gone through the setup process already.
How difficult it is to get running depends on whether you've already got cameras set up and optimized for LPR and are just switching AI engines, or if you mean you're setting the whole thing up from scratch. Camera setup and tuning can be tricky and time-consuming, but there are some threads here that can get you most of the way there. As far as the AI detection and OCR settings in Blue Iris, I didn't find the setup to be difficult at all. It's much easier than what one had to go through to set up external AI tools.
I migrated through several setups in the past. Originally I used DeepStack and Gentle Pumpkin's tool. Over time I switched to CodeProject AI and then Blue Iris's internal AI for the license-plate object identification. Until just the other day I had BI set up to send the images to Plate Recognizer to do the OCR part, but I'm now using the full built-in AI/LPR.
I've always stored the results in an external database using NodeRED as an intermediary to take the alert info from BI and send it to the database. I have dashboards in both NodeRED and Grafana where I can view the reads and images.
TLDR on the rest of this post: I'm now using both BI built-in LPR and Plate Recognizer in tandem, and still find Plate Recognizer to be better at reading tags (YMMV).
-----------------------------
I was having some issues with BI crashing on recent versions, and Ken suggested that I try turning off Plate Recognizer as he doesn't test with it. I've always found PR to be much more accurate in reading tags and didn't want to rely only on the BI OCR, so I spent the past few days making changes to my system that resulted in a rather complicated final setup, but I also was able to add a few new features. I also now can quantify how well the built-in OCR reads tags (TLDR: it's OK but not great).
I now have BI send its LPR results and image to NodeRED. NodeRED then checks the database to see if the Blue Iris OCR results match any tag that has previously been read identically by both BI and Plate Recognizer. If the read matches that list, it accepts the result as valid and does not send it to PR. I have a limited number of PR reads per month, so I don't want to send anything to them for analysis if I don't have to.
If a tag is not on that list, NodeRED sends it to PR for analysis. The query I use to view the data in my dashboard will return the PR tag if it exists, and the BI tag if no PR tag exists. In other words, the read from PR takes priority over the read from BI.
I found that PR provides the vehicle type (sedan, truck, suv, pickup truck, etc) and also what state the tag is from. The state is sometimes 'unknown', but when a value is present, it has been very accurate so far. It also returns a score of how confident it is in the OCR result, which is frequently 1 (meaning 100% confidence). I've added the vehicle type, state, and confidence score to my database.
Plate recognizer also returns multiple tags per image when they are present, something that BI LPR was not doing, even when BI was using PR for OCR. I had to vibe-code some static object handling to avoid saving data for parked cars in my database over and over again, but this now deals with an edge case that sometimes would have missed a read in the past.
I can now approximate how good the built-in OCR is at reading tags (at least in my area, with my cameras). Since PR is highly accurate, if I count all reads where the BI read matched the PR return and divide by the total number of BI reads, I'm getting ~60% agreement over the last 24 hours and ~75% agreement during the past 1 hour. If you have better angles for your cameras or less complicated tags in your area, YMMV.
I have been paying $100 per year to PR for 10,000 reads per month. If BI can read 75% of my tags correctly, I should be able to drop to their free 2,500 per month tier and still get all of my new/unknown tags run through PR.
How difficult it is to get running depends on whether you've already got cameras set up and optimized for LPR and are just switching AI engines, or if you mean you're setting the whole thing up from scratch. Camera setup and tuning can be tricky and time-consuming, but there are some threads here that can get you most of the way there. As far as the AI detection and OCR settings in Blue Iris, I didn't find the setup to be difficult at all. It's much easier than what one had to go through to set up external AI tools.
I migrated through several setups in the past. Originally I used DeepStack and Gentle Pumpkin's tool. Over time I switched to CodeProject AI and then Blue Iris's internal AI for the license-plate object identification. Until just the other day I had BI set up to send the images to Plate Recognizer to do the OCR part, but I'm now using the full built-in AI/LPR.
I've always stored the results in an external database using NodeRED as an intermediary to take the alert info from BI and send it to the database. I have dashboards in both NodeRED and Grafana where I can view the reads and images.
TLDR on the rest of this post: I'm now using both BI built-in LPR and Plate Recognizer in tandem, and still find Plate Recognizer to be better at reading tags (YMMV).
-----------------------------
I was having some issues with BI crashing on recent versions, and Ken suggested that I try turning off Plate Recognizer as he doesn't test with it. I've always found PR to be much more accurate in reading tags and didn't want to rely only on the BI OCR, so I spent the past few days making changes to my system that resulted in a rather complicated final setup, but I also was able to add a few new features. I also now can quantify how well the built-in OCR reads tags (TLDR: it's OK but not great).
I now have BI send its LPR results and image to NodeRED. NodeRED then checks the database to see if the Blue Iris OCR results match any tag that has previously been read identically by both BI and Plate Recognizer. If the read matches that list, it accepts the result as valid and does not send it to PR. I have a limited number of PR reads per month, so I don't want to send anything to them for analysis if I don't have to.
If a tag is not on that list, NodeRED sends it to PR for analysis. The query I use to view the data in my dashboard will return the PR tag if it exists, and the BI tag if no PR tag exists. In other words, the read from PR takes priority over the read from BI.
I found that PR provides the vehicle type (sedan, truck, suv, pickup truck, etc) and also what state the tag is from. The state is sometimes 'unknown', but when a value is present, it has been very accurate so far. It also returns a score of how confident it is in the OCR result, which is frequently 1 (meaning 100% confidence). I've added the vehicle type, state, and confidence score to my database.
Plate recognizer also returns multiple tags per image when they are present, something that BI LPR was not doing, even when BI was using PR for OCR. I had to vibe-code some static object handling to avoid saving data for parked cars in my database over and over again, but this now deals with an edge case that sometimes would have missed a read in the past.
I can now approximate how good the built-in OCR is at reading tags (at least in my area, with my cameras). Since PR is highly accurate, if I count all reads where the BI read matched the PR return and divide by the total number of BI reads, I'm getting ~60% agreement over the last 24 hours and ~75% agreement during the past 1 hour. If you have better angles for your cameras or less complicated tags in your area, YMMV.
I have been paying $100 per year to PR for 10,000 reads per month. If BI can read 75% of my tags correctly, I should be able to drop to their free 2,500 per month tier and still get all of my new/unknown tags run through PR.