Using the machine vision SDK and ATLFlight/ros-examples visual odometry, I am am able to get a visual pose estimate when the unit is stationary, and even when the props spin up to an "armed" speed. However, when the props reach takeoff speed, the state return by the visual pose estimator goes to "failed". I'm guessing this is due to the library's inability to hold onto feature points through frame vibration, but maybe it's something else? Has anyone else seen a problem like this? Qualcomm folk-- any tips on debugging this?
Let me also add that I see this message from the FCU around the time the visual odom fails:
[ERROR] [1493676230.168425876]: FCU: High accelerations, range exceeded 11 times