Aerial robotic tracking of a generalized mobile target employing visual and spatio-temporal dynamic subject perception

This paper proposes a methodology for visual tracking of a dynamic generalized subject within an unknown map, by relying on its perception as a separate entity which can be distinguished spatially and visually from its environment. To this purpose, a 3D-representation of the visible scenery is examined, and the subject is spatially identified by its externally viewed hull via a mesh-connection algorithm aided by visual cues, and visually identified by distinct feature tracking based on an incrementally built list of key-aspects. These two processes operate in closed-loop, and employing a set of assumptions regarding the subject´s structural/temporal invariance the tracking health state can be determined. This work additionally presents the framework for the deployment of this scheme for autonomous aerial robotic subject tracking, employing the dynamic subject/environment distinction to obtain knowledge of the environment structure, and collision-free trajectory generation algorithms to achieve mobile tracking.