Prior Data and Kernel Conditional Random Fields for Obstacle Detection

Abstract

We consider the task of training an obstacle detection (OD) system based on a monocular color camera using minimal supervision. We train it to match the performance of a system that uses a laser rangefinder to estimate the presence of obstacles by size and shape. However, the lack of range data in the image cannot be compensated by the extraction of local features alone. Thus, we investigate contextual techniques based on Conditional Random Fields (CRFs) that can exploit the global context of the image, and we compare them to a conventional learning approach. Furthermore, we describe a procedure for introducing prior data in the OD system to increase its performance in familiar terrains. Finally, we perform experiments using sequences of images taken from a vehicle for autonomous vehicle navigation applications.