Vision-aided inertial navigation using three-view geometry

This paper presents a novel unscented Kalman filter based algorithm for vision-aided inertial navigation system (VINS). It uses dynamic model of inertial measurement unit (IMU) to perform state propagation and trifocal tensor based geometric constraints of three views to update system. Unlike the conventional methods, the positions of feature points are neither required to be augmented into system state, nor estimated during initialization. The main contribution of this paper is twofold. First, a dynamic model which considers three-view geometry is derived for three-view based VINS. Second, it is the first time that trifocal tensor based geometric constraints and point transfer of three-view geometry are used for VINS, gaining robustness and avoiding scale ambiguity. The approach is experimentally evaluated by using a real IMU and image dataset that was recorded by a ground vehicle, verifying its effectiveness.