ON THE OBSERVABILITY IN VISUAL SLAM NETWORKS

Observability is an essential aspect for the performance of a visual Simultaneous Localization and Mapping (SLAM) network. This paper presents a statistical perspective to evaluate the observability of visual SLAM networks and its dependence on network structure. In particular, we first give the general form for the Fisher information matrix (FIM) of each visual observation and the impact of 3D point translation and camera motion on observations. Then the observability of visual SLAM networks is investigated based on the nullspace of the FIM and its relation with network structure to derive the lost rank and its upper and lower bounds. We also propose the ill-conditioned score to evaluate the degradation of visual SLAM performance under any network structure. Numerical results validate our conclusions.