Prior Data and Kernel Conditional Random Fields for Obstacle Detection

Obstacle detection (OD) is important in many mobile robot applications and autonomous vehicles. The most successful OD systems rely on range information to detect obstacles by size and shape. The most popular and widely used sensors in this area are laser rangefinders, due to their reliability and quality of data. Unfortunately, they are expensive for some applications, such as the automation of tractors in agriculture. For this reason, in this work we train a OD system based on a monocular color camera, which is comparatively inexpensive. The lack of range data is compensated by exploiting the contextual information present in the image. This paper investigates contextual techniques based on Conditional Random Fields (CRFs) and compares 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.