EPVC: a novel initialization approach of visual-inertial integrated navigation

The fusion of visual and inertial measurements in robotics community is growing in popularity since both of them have complementary perceptual information. Pre-initializing gyroscope bias and accelerometer bias of the inertial measurement unit (IMU) is a critical issue to achieve a better fusion performance, and the metric scale is another crucial element to be estimated. Current mainstream loosely-coupled initialization methods are unstable as they do not incorporate IMU information into the visual structure from motion. In addition, the accuracy of the tightly-coupled methods is limited since they do not use visual observations to compensate gyroscope bias and usually ignore them in close-form solution. In this paper, a visual-inertial (VI) initialization method which we refer to as epipolar plane normal vectors coplanarity constraint (EPVC) method is proposed to solve gyroscope bias. A step further, a novel analytical solution is presented to optimize other parameters. Comparing the proposed method with VI navigation systems-mono and inertial-only optimization through the publicly available EuRoC dataset, the results demonstrate that the proposed method outperforms the existing methods in estimating the gyroscope bias and scale, and with the increase of initialization time, the accelerometer bias error and gravity direction error have a clear diminishing tendency.