Monocular heading estimation in non-stationary urban environment

摘要

Estimating heading information reliably from visual cues only is an important goal in human navigation research as well as in application areas ranging from robotics to automotive safety. The focus of expansion (FoE) is deemed to be important for this task. Yet, dynamic and unstructured environments like urban areas still pose an algorithmic challenge. We extend a robust learning framework that operates on optical flow and has at center stage a continuous Latent Variable Model (LVM) [1]. It accounts for missing measurements, erroneous correspondences and independent outlier motion in the visual field of view. The approach bypasses classical camera calibration through learning stages, that only require monocular video footage and corresponding platform motion information. To estimate the FoE we present both a numerical method acting on inferred optical flow fields and regression mapping, e.g. Gaussian-Process regression. We also present results for mapping to velocity, yaw, and even pitch and roll. Performance is demonstrated for car data recorded in non-stationary, urban environments.