ANPR feature request

Post Reply
JBrad
Posts: 3
Joined: Tue Aug 10, 2021 6:00 pm

ANPR feature request

Post by JBrad » Tue Aug 17, 2021 10:53 am

Now that BI's Deepstack integration is fairly mature & robust I've fully switched to it over the fantastic AITools.
And since I can now reliably detect vehicles over people, cats, dogs, etc I've finally been giving ANPR a try.

---

One feature I'd like to request is one which masks or crops anything outside the detected object bounding box(es) before sending the image to be analysed for ANPR.
...Maybe do a simple crop if only one object detected, mask non object rects and crop to overall boundary if multiple.


This would be a useful addition for a number of reasons...

* Would cut down on detections of stationary parked vehicles in frame
* Would cut down on instances of camera timestamps being confused for reg-plates
* Would mean less bandwidth usage uploading images to be processed (eg. 0.5MP image just of the vehicle being queried vs a 4/8MP image of the entire scene)
* Would mean faster uploads for those of us with crappy rural broadband connections
* Would be far faster and less CPU intensive for the ANPR service to process the image and return a result

The last two are probably the most pertinent, having a confirmed alert back faster is always a good thing.

---

Also, Something I've already made a rudimentary attempt at using Home Assistant but would be nice to have integrated would be the ability to supply BI a whitelist of Reg plates (assigned into groups) and then BI uses some fuzzy logic to determine if a known reg plate was detected then trigger an action (or not dependent on what is detected). Eg. don't flag an alert if one of my own vehicles triggered the motion, fire a specific alert if a plate in the 'delivery services' group, fire a different one if in the 'friends/family' group, etc.

The fuzzy logic would score how close of a match a plate is to one on the whitelist allowing some level of error in the plate matching. Allowing for substitutions (8-B, W-Y, 0-O), missing letters, letters added, etc. ...you'd set a threshold for how much error is allowed.

I know we have the 'myplates' box, but this would take that feature a step further.

My current DIY Node-Red / Home Assist method doesn't do any of the match scoring (requires an exact match), and doesn't specify groups, just a simple white/black list (no alert, special alert, 'unknown' alert). Was going to add this to my implementation, but thought I'd hang fire in case it's something which is a planned feature and I'd be wasting my time.
Post Reply