Erroneous data in VOA distance bin/linescan

[Kavin Kailash]

Hello ModalAI Dev Team,

@Alex-Gardner @Chad-Sweet

We appreciate the great work done by the ModalAI team and look forward to more amazing feature rollouts.

While working on setting up the VOA utility packaged with the VVPX4 service, we noticed the presence of ‘ghost’ obstacles facing the front of the drone. We noticed this while inspecting the VOA linescan data from VVPX4 by running the following command:
voxl-vision-px4 -s
These erroneous data points seem to fluctuate around 0.25-0.30m and don't seem to respond to presence of any actual obstacle. We have double checked that the front stereo sensors are not obstructed by the airframe. Even while inspecting the OBSTACLE_DISTANCE MAVLINK message in QGC, we see the same behaviour in the data points.

In contrast, the data from the rear section of the VOA bin seem to correlate with our handheld obstacle tests. We attempted to flip the cameras from back to front (keeping in mind the R’s and L’s) and swapped the calibration (intrinsics) as well. But we still observed the same behaviour on the front sector of the VOA distance bin/linescan. It should also be noted that both the stereo pairs were re-calibrated recently.

We would be sharing more information regarding this as we collect more data, but in the meantime we would greatly appreciate it if you could assist us in debugging this behaviour.

HW Platform: VOXL2 with Front/Rear Stereo, HiRes and Tracking
System Image: VOXL2 Platform 0.9

Thanks,
Kavin

[Chad Sweet]

I'm assuming you have calibrated the stereo sensors?

When you view the DFS output, do you see ghost objects appear?

[Kavin Kailash]

@Chad-Sweet Thanks for your response. I apologise for the late response. As you correctly pointed out, we resolved the issue after a re-calibration of the front stereo pair in natural lighting and by further tuning the 'blur_size' and 'post_median_size' parameters in the DFS server config file. Now the data reflects actual physical obstacles quite reliably.