6. Conclusions
In this paper, we presented the UDQ-based graphical SLAM approach,
in which vertices and edges are represented with UDQs,and the relative transformation between any two vertices can be
illustrated by a UDQ matrix. Compared with the HTM constructed
by a rotation matrix and a 3D translation vector, the UDQ matrix
is specified by its UDQ elements without trigonometric computations.
We formulated measurement errors with the UDQ product,
which obtains a linear form with respect to each pose UDQ. Subsequently,
the Jacobian matrices of the error function can be efficiently
computed by the product of corresponding UDQ matrices,
which have been evaluated in the computation of measurement
errors. Since the linearized error function has a same structure as
state-of-the-art approaches, the Gauss–Newton was directly applied
to solve this optimization problem. In addition, we compared
two popular over-parameterized ways, namely the quaternion and
the axis–angle, in the graph-based SLAM optimization problem.
Under small perturbations, the optimization method based on either
one can obtain same results. However, the quaternion-based
method results in a slow converge speed in large noise cases, due
to its numerical instabilities in the recovery from the minimal
representation to the full representation. Finally, we provided extensive
experimental results using publicly available and simulated
datasets in 2D and 3D environments to verify the feasibility
and efficiency of the proposed method, and proved that the overparameterized
ways based on the scaled axis–angle outperforms
the one based on the quaternion.