Optimization Based Mono-Visual Inertial Odometry

Visual-inertial odometry has exerted a variety of promising applications in the robotics and computer vision society due to their complementary nature in terms of both accuracy and quick response. This paper presents a novel odometry approach using monocular camera measurements and inertial cues. It is formulated within a probabilistic framework that considers both visual landmark reprojection errors and inertial statistical properties. Unlike the traditional optimization based methods, ours takes the inertial cues in the propagation step, rather than incorporates them in the cost function, with the aim of reducing the system computational burdens. The inertial estimates are then fed into the cost function pose part and visual feature matching part. Additionally, to bootstrap the camera sparsity and shorten the computational time, merely the keyframe lankmarks are selected and the less informative states are marginalized in the objective function. The experiments show that our approach presents competitive outcomes in terms of both accuracy and robustness