Vision-Based Obstacle Avoidance Using Active Scene Segmentation

Successful operation of autonomous flight systems in uncertain environments is currently limited by the lack of practical obstacle avoidance systems. In the proposed effort, we shall exploit recent and on-going advances in the fields of image processing, estimation, real-time path planning, and guidance/control to demonstrate successful flight through a cluttered and uncertain 3-D urban environment using simple imaging sensors and appropriate custom-developed processing hardware. In particular, we propose the innovative use of variational methods to dynamically segment scenes, leading to a fast and natural approach to estimating the location of unknown 3-D obstacles. During phase I the feasibility of real-time algorithm execution will be established using current generation processors, and robustness to transient sensor data, distortion, and obscuration evaluated. A systems engineering analysis will be used to fully characterize system performance as it relates to a variety of design parameters and constraints. In phase II, efficient means to produce the appropriate level of image understanding required for successful operations will be further developed, the algorithms will be fully realized in hardware and software, and the prototype system will be fully evaluated in UAV flight operations over urban terrain.