For this summer I have been interning with Motion Picture Marine which designs and builds the Perfect Horizon. The prime use of the Perfect Horizon is to keep a camera level with the Horizon. This is especially important when working on boats and shooting off into the distance. In the near future it is also expected that keeping a camera shot level with the horizon will become even more important because 3D moves on average have shots that are twice as long, and any movement in the image is many times more likely to cause motion sickness of the audience.
The way that the Perfect Horizon works is by compensating by movements of the boat or moving vehicle using two motors, to move the platform in the opposite direction. The problem comes from when the Perfect Horizon is not mounted on a stable platform. This is because when a 100 pound camera is mounted on top, there is a large deal of inertia, and thus the platform underneath ends up getting pushed. If the platform can be moved even a little bit, then the Perfect Horizon will detect this and attempt to compensate as it was designed to do. In turn this causes the system to begin to vibrate. (This problem can easily be seen on a crane arm, or any poorly designed mount)
This problem can basically be broken down into two major components. The first being the detection of the vibration, without miss detecting normal operation. And the second is modifying the internal parameters of the system to prevent it from vibrating.
For my first attempt at detecting the vibration I just attempted to use simple signal processing filters. This ended up working on my test rig as it was a stationary platform. In moving to a moving platform I began to notice some problems with the solution. The first being that it would miss detect jolts or changes in motion as vibration. This was simply unacceptable as this would degrade the quality of the system and prevent it from doing the job that it was designed for. With some internet research, I found out that jolts are visible on all frequencies while vibrations are only on a select few. (link) This lead to using fourier transform for processing the signal. What a fourier transform basically does is it converts signals in the time domain into signals in the frequency domain. This basically means that when looking for a vibration, we need only check for outliers in the set of data. This has turned out to be a better solution, and combining it with simple check to ensure that there are enough samples that indicate vibration there has not been any false positives as of thus far.
The second problem came in with modifying the gain of the system to prevent it from vibrating. I believe that part of the problem with the vibration comes from the fact that the system produces thousands of tiny jolts while moving, this is because the system will determine where it wants to move to, and then it will move to that point up to a maximum speed. Once it has reached that point, it will stop and wait until the next sample is ready. The solution here is to slow down the motors so that they do not reach the end before the next sensor reading. The obvious answer of just always slowing down the motors will not work, as then the system would not be able to deal with the larger movements because the motors would be moving too slow. In implementing the solution the motors are only slowed down when the system determines that it is vibrating. This essentially gives the system some form of adaptive control over the internal parameter used in the system.