voxl-mapping position drifting
playing with the voxl-mapper I was getting weird results where it seemed like the drone position was drifting through space, creating long "tunnels" in the mapping or just broken geometry. However the QVIO overlay looks fine and inspect-qvio mostly shows OK status. Fisheye lens was calibrated with 35mm chequerboard.
What causes this drifting?
Since I had more success in the sunny morning (indoors), how sensitive is QVIO to light conditions?
One of the biggest outstanding issues with QVIO is blowup detection, right now the status of the algorithm almost always says ok even if in reality it's drifted off into oblivion. Does the x/y/z position of the drone look correct in the overlay or does it say that it's flying off? these issues arise because the qvio algorithm is closed source (even to us) and we've just written a wrapper on top of it. We're actively working on getting a couple different open source vio algorithms working that will allow us to dig deeper into their behavior and better detect blowups, but this is still a little while off being as robust as qvio when it's working.
I also noticed the scaling of the QVIO coordinates is wrong. I had to recalibrate the sensors because I had to reflash the drone. Maybe you can pull the factory calibration files for my device because I suspect my calib is bad. Serial is M2100000CCS
Steve Turner last edited by
@Alex-Gardner Definitely interested in open VIO options. Would be awesome being able to use other techniques for picking out key points and even accelerate them on the hardware. Maybe QVIO already does that...