YARF: An Open-Ended Framework for Robot Road Following

摘要

This thesis describes YARF (Yet Another Road Follower), a vision- based system for autonomous road following. Video data from a camera mounted on a robot vehicle is fed into a computer, which analyzes the image data to locate the position of the vehicle relative to the road. The computer then issues command to actuators attached to the throttle, brakes, and steering in order to drive the vehicle along the road. YARF has been extensively tested using a combination of open- and closed-loop runs on testbed vehicles, simulation, and data from videotapes. YARF provides a set of perception capabilities to locate the position of the vehicle relative to the road, to detect changes in the lane structure of the road, to navigate through intersections given a model of the intersection geometry, and to extract the lane structure of the road without a prior model. The central theme of YARF is that using richer models improves road following performance. Models of geometric structure, of road appearance, and of segmentation performance all simplify processing and contribute to improve reliability. YARF uses road models in several ways: Model drive segmentation; Exploitation of model coherence to avoid the influence of contaminating data; and Data driven recognition of model changes. While YARF assumes that the models of road structure used are generated off line, the thesis presents an algorithm designed to automatically extract much of the needed model information. The algorithm uses a weak domain model to filter a noisy image segmentation, extracting both feature geometry and type.