Motion Estimation for Obstacle Detection and Avoidance Using a Single Camera for UAVs/Robots

Abstract

This work focuses on efficient and effective vision systems for object detection for ground & aerial robots venturing into unknown environments, with minimum possible vision aids onboard, i.e. a single camera. The existing approaches to solve such a problem include ‘Structures From Motion’, Optical Flows and ‘Flow Field Divergence’, etc. These approaches have been analyzed here for various constraints involved, and the Motion Estimation technique has been proposed to solve for obstacle detection problem for collision avoidance, using a single camera. This technique not only overcomes various constraints of other approaches, but also retains most of their merits. Its implementation on synthetically generated images as well as on some real videos from a UAV, has been proved successful to solve this problem. However, while implementing the technique on UAVs in actual flight, quite a few undesirable motion vectors have been encountered. An analysis of such implementation issues has been presented next, along with proposed solutions. After adequately addressing such issues, the application of this approach is being attempted on any of Georgia Tech’s aerial robotic platforms.