Using computer vision (monocular odometry) to correct vehicle path

Hi there,

I have an idea - you could correct GPS vehicle paths with some simple hardware and software - use a $10 OBD2 dongle to get vehicle speed and than some simple computer vision (monocular visual odometry) to correct the GPS vehicle paths that are often very erratic especially within city down towns. The vision code is few lines only (like here for example) but could bring huge gains to path accuracy. The idea is use GPS for long distance correction and otherwise rely on the visual odometry.

With the setup above you may actually be even able to collect quite precise vehicle paths and use them for precise mapping. Mobileye company does something similar with their REM (Road Experience Management) technology for autonomous driving. They also collect sparse 3D data to use as “checkpoints” in their map - well this principle is very simple and I think it cold also be done within a mobile app.

Have you been thinking about doing something like this?


They know about it, maybe not about the monocular thing.
I have two OBD2s and none works well.
There is already enough misery.

Could this possibly work where there is no GPS such as tunnels and underground?


Using OBD2 for reading speed will be quite reliable I think. All cars provide vehicle velocity in OBD2 standard since almost twenty years. There are also affordable OBD dongles that work well (you just can not use the very cheap chinese ones even though many of them work flawlessly).


If saved photos were used as stationary checkpoints then yes it could be used for navigation even without GPS. It is a bit more complicated but in general it is like I say with little computer vision “magic” behind it.

Mapillary’s competitor, I have forgotten its name, uses OBD2.

Yes, they are called OpenStreetCam.