Probabilistic egomotion from a statistical framework
Harshil Shah, A. Lakshmikumar · 2007
Traditional egomotion estimation algorithms have largely depended on deterministic feature correspondences to infer information about the camera and have been oblivious to the scene geometry by treating scenes with varying projectivities uniformly. This paper builds on the statistical framework of the joint feature distribution (JFD) which models the joint probability distributions of the positions of corresponding features in different images. This framework explicitly gives probabilistic correspondence search regions that can be stably estimated for the whole range of planar, shallow and deep scenes using relatively few correspondences. These joint probability distributions are constrained by the epipolar constraint to yield a distribution over all feasible egomotions. The paper also compares the proposed method against existing well-known methods and quantifies the improvements in the egomotion estimates.