Fusion of visual odometry and inertial sensors using dual quaternions and stochastic cloning

The article presents an approach where the visual odometry from the stereo camera is fused with the inertial measurements using an error-state extended Kalman filter. A simplified strapdown inertial mechanization is used to predict the orientation and position of the user while the measurements of the visual odometry are employed to reduce the fast position drift of a pure inertial approach. Within the work the relative orientation and position measurements are considered as a generalized displacement represented as a dual quaternion. Using the framework of stochastic cloning both the rotation and translation measurements are incorporated in a statistically consistent way including a proper treatment of the uncertainties of the previous estimates. An enhanced performance compared to classical dynamical models is obtained for a basic sensor hardware setup and demonstrated for several typical experimental data sets. The described sensor system is completely self-contained and can be used standalone for indoor or combined indoor/outdoor scenarios as well as in hybrid positioning systems.