Fail-safe visual-inertial navigation for UAVs

In this paper, we propose a visual-inertial state estimation framework which is able to detect and mitigate failure modes to ensure best possible state estimation for platform control at all times. The main focus here is on the proposed sensor switching method which allows seamless switching between integration of pure inertial cues, the use of inertial-optical flow based velocity estimates, and the use of visual-inertial based position estimates for the control of an inherently unstable aerial vehicle. The switching mechanism automatically detects if a state estimator part is faulty and reduces the sensory input to the remaining, healthy, information streams. In addition, a re-initialization sequence is run for the faulty segment until the full system is recovered. With the additional capability of each segment for self-calibration, the system is both self-calibrating and self-healing. The full framework has been integrated on an embedded platform on-board a real 500g small aerial vehicle and run at 30Hz camera stream and 1kHz inertial readings for live demonstration.