Sensors Calibration and Filter Initialization in Visual Inertial Odometry

Visual inertial odometry (VIO) is an enabling technology for mobile home robots which is possible nowadays due to novel low-cost cameras and inertial measurement units (IMUs). This work analyzes how the system-hardware calibration and system-state initialization, followed by the dynamical propagation of IMU readings, involving intrinsic bias, influence the VIO. Then we analyze the synergies between these two sensors in the context of pose estimation. The calibration of the system, namely the rigid transformation between camera and IMU is performed with a free, public domain, tool. The visual and inertial data is processed by means of an Unscented Kalman Filter with Lie group embedding for state representation. We propose a state initialization for this filter that enables matching the integrated IMU readings with the tracked visual features. Experiments and results show the importance of successfully matching the feature tracks in the first images with the starting integration of IMU readings, as significant errors in initial bias estimations may preclude sensors fusion and filter convergence. Results show also that the operation of the fusion filter allows synergies between the two sensors, while the IMU provides instantaneous and reliable estimations of translation and rotation speeds, the visual component provides IMU bias correction.