More AHRS thoughts

I had a moment of panic when I thought about deconvoluting the accelerometer's measurement of tilt (a positive nose-up pitch θ adds a positive gsinθ in the x-axis accelerometer reading) from its measurement of translational acceleration (a positive forward acceleration also adds a positive a_x into the x-axis accelerometer). The scary part is that positive pitch requires forward cyclic to return to hover, whereas forward acceleration requires aft cyclic to return to hover, so distinguishing between tilt and translational acceleration will be important to the control system.

In the short run, we can deconvolute the two by integrating the gyro signal to get attitude, but gyro drift makes this infeasible in the long run. In the long long run, our plant tells us we can have a steady-state tilt but not a steady-state acceleration (unless we're flying in a circle...hmm...), so we might be able to deconvolute the two that way. I'd need a good plant model in the estimator, though. Even so, I think it'll all hinge on the gyro drift.

What about other methods of attitude sensing? If we can solve attitude sensing, running optical flow on a downward-pointed camera can tell us a lot about velocity and altitude. Maybe more cameras? On the sides, for horizon sensing? Infrared horizon sensing? Two downward-pointed cameras to generate a parallax field, assuming a flat ground? What if the surface is too low-contrast (like grass)? Or not flat (like a hillside or building)? How much can the magnetometer tell us? What about adding accelerometers? This particular problem is a fundamentally about perceived acceleration in a non-inertial frame, but adding another accelerometer at a different location on the helicopter would give me angular acceleration, which could improve the gyro readings...

One last thought: do we try to integrate the human more? Maybe the goal here is not to strictly stabilize, but to slow down the dynamics to the point where the operator can take over these difficult sensing problems. In this case, (assuming the goal is a hover) the operator would distinguish between forward acceleration and nose-up pitch by his choice of control input.