First year report

摘要

An autonomous robot surviving in the 3D world may sample its environment as a 2D image sequence, each image differing slightly from its predecessor as a result of robot/scene motion. Furthermore, such a system may not require specific identification of objects within a scene but rather be more concerned as to which objects pose a threat or obstacle to some future action, or to the recovery of information about the robot's own motion relative to the scene (the so called ego-motion). This document introduces the concept of 3D environmental structure and sensor motion recovery from 2D image sequences and more especially the concept of an optical flow field. Problems inherent in both sensor and optical flow field characteristics are described and their consequences discussed. By considering the informational content of variously derived flow fields the requirement for 3D environmental structure constraints is introduced. Specific algorithms are classified, detailed and reduced to a canonical format. A parametric noise model is proposed and accuracy bounds derived. Conclusions and further work are presented.