Long Term Development - Vehicle Recogntion


Hi all,

I am interested in increasing read rates overall so I can get a high match rate. from entry point to exit point (2 reads)

I want to know if anybody has thought of addiitonal identifiers like vehicle shape and colour?




Hi Glenn,

Yes, that is a great idea. The problem with using make/model/color right now is that it’s a bit of a binary match. For example, if it classifies the body type incorrectly, it won’t match. Or if it sees the make emblem clearly from one angle but not the other, it may not match.

One of the things we plan to add when we add general vehicle detection later this year is a vehicle similarity score. We need to implement this in order to track vehicles as they drive across a camera. That would be more appropriate for something like this because it would give you confidence that the car is the same, even if the make/model/color/body type doesn’t match exactly. Perhaps that is something we could expose in the API (e.g., given two images, compute the vehicle similarity).


That is very encouraging. I also think that there is value in the ‘matching’ scores. Not just the ability to read a plate correctly.
For me I would like to see ways to match vehicles regardless of how many characters were correctly read. So if there were say 4 characters read well and a couple incorrectly read (say O, 0 etc…), then at least it can match car on colour.
Also capture rates. If the system can successfully identify the locaton of the plate, then this is like finding the nose on a person’s face using facial recognition, right. Now you have a strong identifier as the position of the plate relative to the overall object would be a huge differentiator.