Probabilistic Obstacle Partitioning of Monocular Video for Autonomous Vehicles

This paper reports on visual obstacle detection from a monocular camera for autonomous vehicles. By leveraging a textured prior map, we propose a probabilistic formulation for finding the optimal image partition that separates obstacles from groundplane. Our key insight is the use of a prior map that enables ground appearance models conditioned on prior map texture and a probabilistic optical flow vector formulation derived from known scene structure and camera egomotion. We evaluate our methods on a challenging urban setting using data collected on our autonomous platform and we demonstrate that a notion of obstacles in the camera frame can improve visual localization quality.