Autonomous Obstacle Avoidance for UAV based on Point Cloud

This paper presents an autonomous obstacle detection, avoidance, and path-finding algorithm for UAV (un-manned aerial vehicle) in outdoor applications by using a stereo depth camera. The obstacle avoidance part is divided into two steps: process the point cloud, remove noise and points of ground, and extract clusters to represent different objects; perform and define detectors, to detect whether the UAV has a collision at different distances and in various directions around the nearest obstacle, even behind the free escape space respectively. Finally, the flight decision is selected with the direction with highest votes and publish the flight command to be executed by UAV. A voting mechanism for judgments during a period is implemented, to avoid the misjudgment by a single frame, as well as to enhance the plausibility of flight decision. The algorithm is evaluated in Gazebo-based simulation environment as well as in field tests, has been approved that the UAV can successfully avoid obstacles in dense woods and find new obstacle-free paths, even in extreme environments with huge buildings, which indicated that the algorithm achieved its initial purpose.