Stratégies de guidage visuel bio-inspirées : application à la stabilisation visuelle d’un micro-drone et à la poursuite de cibles

• Manecy Augustin

• Quadrotor
• Hovering flight
• Optical sensor
• Attitude estimation
• Bio inspiration
• Gaze stabilization
• Hyperacuity
• Insect
• Vision
• Vestibulo ocular reflex
• Visual fixation
• Accurate tracking
• Modeling
• Linear control
• Nonlinear control
• Parameter identification
• State observer

Insects, like hoverflies are able of outstanding performances to avoid obstacles, reject disturbances and hover or track a target with great accuracy. These means that fast sensory motor reflexes are at work, even if they are minimalist, they are perfectly optimized for the flapping flight at insect scale. Additional refined mechanisms, like gaze stabilization relative to the body, allow to increase their flight capacity.In this PhD thesis, we present the design of a quadrotor, which is highly similar to an insect in terms of perception (visual system) and implements a bio-inspired gaze control system through the mechanical decoupling between the body and the visual system. The design of the quadrotor (open-source), itspilot and its decoupled eye are thoroughly detailed. New visual processing algorithms make it possible to faithfully track a moving target, in spite of a very limited number of pixels (only 24 pixels). Using this efficient gaze stabilization, we developed new strategies to stabilize the robot above a target and finely control its position relative to the target. These new strategies do not need classical aeronautic sensors like accelerometers and magnetometers. As a result, the quadrotor is able to take off, move and land automatically using only its embedded rate-gyros, its insect-like eye, and an altitude measurement. All these experiments were validated in a flying arena equipped with a VICON system. Finally, we describe a new toolbox, called RT-MaG toolbox, which generate automatically a real-time standalone application for Linux systems from a Matlab/Simulink model (http://www.gipsalab.fr/projet/RT-MaG/). These make it possible to simulate, design control laws and monitor the robot’s flight in real-time using only Matlab/Simulink. As a result, the "time-to-flight" is considerably reduced and the final application is highly reconfigurable (real-time monitoring, parameter tuning, etc.).