Feature Detection Degrading Resulting In Localization Failure

[jake3991]

I'm running this pipeline in ROS using the instructions provided. I used the ROS drivers opencv video node to send a live feed from a USB camera to the svo_ros system. The good news is it works! But I'm having an issue with the system.

When I start to move the camera (slowly) the number of features degrades from 150 to the low 50s high 40s which results in the following messages.

[ WARN] [1516128738.966920576]: Tracking less than 50 features!
[ WARN] [1516128738.966977358]: Lost 46 features!
[ WARN] [1516128739.000414726]: Relocalizing frame
[ WARN] [1516129068.547778792]: Not enough matched features.

It's almost like either the FAST parameters or refresh parameters are wrong. Does anyone have any idea what could be wrong here? I'm mainly confused because the demo video for this system depicts a camera being moved very quickly.

[yan99033]

Been playing around with SVO a bit, the main reason that the number of features dropped significantly is that depth-filter is not able to converge the newly initialized seeds through triangulation of corresponding features in the subsequent frames. This failure mode is mainly due to having a forward looking camera and fast camera motion.

If you dig deeper, each feature in the keyframe is initialized with averaged scene depth, which works remarkably well if you have a planar scene. In addition, you will also need a higher frame-rate camera so that the true depth can be obtain statistically through Gaussian X Beta model.