Noise model creation for visual odometry with neural-fuzzy model

In this paper, we propose a technique of learning a noise pattern of visual odometry for accurate and consistent 6DOF localization. The noise model is represented by three parameters of feature points as input: (I) The number of inliers among feature points, (II) Average of distances between feature points, (III) Variance of interior angles. The correlation between these parameters and estimation error is also described. To approximate the complicate noise model accurately, our technique adopts Hybrid neural Fuzzy Inference System (HyFIS) for a learning engine. The noise model is created with HyFIS beforehand, and then the error of visual odometry is estimated by the noise model and compensated on the fly. Learning results of the noise model and results of 6DOF localization in untextured and dynamic environments are presented, effectiveness of our technique is shown.