Visual collision detection for a mobile robot

This thesis describes a new framework for collision detection. The key result is a novel parametrization of the motion space from which collision information can be directly inferred. In the development of our approach, we present an original exposition of the looming detection problem, in which we extend the notion of focus of expansion (FOE). The parametrization procedure involves a prior velocity estimation stage, for which we propose a new 4-D plane-fitting algorithm for calculating the normal components of velocity. This method is not only simple and direct, but also provides for a reliability measure to evaluate the goodness of the velocity estimates. These estimates are then combined using voting, which is a technique robust to noise, before being built into a cost function. To the costs function we apply an iterative optimization process, which yields the optimal parameter-space triple $(v sb{x},v sb{y},{ cal R}).$ From these parameters, it can be determined whether or not an object in the scene will collide with the observer, and if so, the time of contact. The thesis also describes the implementation of this new approach. The results are encouraging, and indicate that the method is robust to both noise and occlusion, and performs well in the presence of multiple objects