Vision-Based Obstacle Avoidance Using Active Scene Segmentation

A program is proposed to develop, flight demonstrate and transition a robust capability to autonomously detect and avoid obstacles and features typically encountered in three-dimensional flight by relatively small fixed or rotary wing unmanned air vehicles in low-altitude urban operations. There are two basic parts to the problem: (1) continually detecting and modeling the observable 3-D obstacle field in real-time, and (2) autonomously guiding the vehicle to accomplish a given set of mission objectives while ensuring there are no collisions with obstacles. The key innovation of the proposed approach is development of a technique to accomplish the stated objective that requires only a single 2-D image stream (available on essentially every small unmanned air vehicle in the inventory today). Demonstrated phase I algorithms will immediately move into practical implementation and evaluation on an existing low-cost flight test vehicle. Meanwhile, hardware and software design specific to integration on the Puma small unmanned air vehicle will be completed. Option 1 will demonstrate improvements in Puma operational effectiveness using the developed technology in simulation, and complete fabrication and development of Puma specific hardware and software. A second option will provide for technology demonstration on a Puma flight system in collaboration with Aerovironment, Inc.