I am working on an estimation application for multiple robots, each of which uses measurements from various sources to calculate position and orientation data. For now, I am looking at about three sources:
- Orientation from an IMU
- Both position and orientation from a camera
- Position and orientation from relative measurements w.r.t another robot.
Naturally, the most widely used technique that combines a model with measurements is something like an EKF. But in my case, the orientation data (IMU) comes from an autopilot, which already has an EKF onboard, and hence provides covariance estimates as well. The vision based estimation (both individual and relative) are computed through a few iterations of bundle adjustment, and the bundle adjustment solver also provides covariance of its resultant estimate. Finally, I am not really interested in utilizing a complex, non linear model for the robot, but mostly in just fusing the measurements to provide one final pose estimate.
I have read about the concept of 'covariance intersection' in the context of Kalman filtering, which has been implemented in cooperative pose estimation using multiple data sources. I was wondering if I could get some advice as to whether Kalman filters etc. would still be applicable in my case, and if so, how to adapt them?