Robust Person Following with an Autonomous Quadrocopter using a Depth Camera

(3D Bild-Analyse und Synthese)
Externe Arbeit
Betreuer: Anas Al-Nuaimi
Person detection and following is an integral part of an environment where robots interact with humans. This thesis deals with robustly following a person with a quadrocopter using a depth camera. A quadrocotper is an UAV (unmanned aerial vehicle) that is lifted and propelled by four rotors. Various algorithms have been developed for person following by UGV (unmanned ground vehicles). Tracking humans with UAVs is relatively a novel problem. Aerial vehicles make person following a difficult problem due to their complex flight dynamics. Existing work demonstrate marker based motion capture systems to use human pose for control. In this thesis a framework for following a person using marker-less motion capture will be developed. Most existing trackers make the assumption of a static background which is clearly violated on flying robots. Major contribution of this work is to investigate how a motion estimate from visual odometry and IMU can be used to stabilize the camera images such that existing trackers become applicable. Compensating for the camera motion in subsequent frames creates a video from a virtual static camera. It allows to track a person in the stabilized video. The estimated pose will be used for control, i.e., to actually follow a person. Many potential applications exist for this system. One is filming, as it allows to record a video of a moving person from a third-person perspective without an additional cameraman. External visual imagery is used by athletes to monitor their activities while training. A quadrocopter following an athlete and providing a video which can also be used later for detailed offline analysis is also an exciting application of such a system.
Keywords: Computer Vision