Core Concepts
The unobservable directions of Vision-aided Inertial Navigation System (VINS) are uniform global translation and global rotations about the gravity vector, while the unobservable directions of Lidar-aided Inertial Navigation System (LINS) are the same as VINS, requiring only one feature to be observed.
Abstract
This paper analyzes the unobservable directions of the nonlinear models for Vision-aided Inertial Navigation System (VINS) and Lidar-aided Inertial Navigation System (LINS).
For VINS, under the assumption that there exist two features observed by the camera without occlusion, the unobservable directions are:
Uniform global translation
Global rotations about the gravity vector
For LINS, the unobservable directions are the same as VINS, but only one feature needs to be observed.
The analysis is done by calculating the Lie derivatives of the observation functions and determining the null space of the observability matrix. Key results include:
When there are two features that are linearly independent, the null space of the observability matrix has a specific structure, with the unobservable directions corresponding to global translation, global rotation about the gravity vector, and the bias of the gyroscope.
The rank of the observability matrix is shown to be 3N+11, where N is the number of features.
The unobservable directions are characterized in a compact form, enabling better understanding and utilization in practical VINS and LINS implementations.