In this paper, we focus on the problem of online intrinsic camera parameter calibration for a visual-inertial system. Imprecise intrinsic camera parameters will result in unreliable pose estimation or even cause estimator divergence. Specifically, we present a nonlinear observability analysis of the system and prove that there are four unobservable directions spanning the right nullspace of the observability matrix, i.e., the rotation about the gravity vector and the positions in the global frame. We propose an extended Kalman filter-based visual-inertial odometry method for calibrating intrinsic camera parameters while estimating the pose simultaneously. The observability properties and the performance of the estimator are validated using both the simulated and real-world datasets.
All Science Journal Classification (ASJC) codes
- Electrical and Electronic Engineering