Visual Control of a Miniature Quad-Rotor Helicopter

Automatically controlling the flight of an unmanned aerial vehicle has typically been achieved using a suite of sensors to give reliable estimates of the vehicle’s position, orientation and velocity. Recently, on-board video cameras have been used as a primary source of these estimates, for the outdoor flight of large model helicopters. Here a much smaller, indoor model helicopter is automatically flown by employing just a single miniature transmitting camera. However, this task presents a stiff set of challenges to the visual tracking system used to obtain pose and velocity estimates from the transmitted images. This thesis shows how they can be overcome. Tracking discontinuous motion using ambiguous image features means that it is not sufficient to consider just a single hypothesis for the camera pose. A method for efficiently generating a representation of a multi-modal posterior probability distribution is presented. The technique combines ideas from ransac and particle filtering such that the visual tracking problem can be partitioned into two levels, while maintaining multiple hypotheses throughout. The resulting system demonstrates a significant improvement in tracking reliability over previous unimodal approaches. The use of multiple hypotheses however greatly increases the computational complexity of tracking and this causes difficulties when operating in real-time. This is addressed using a technique for clustering measurements which simplifies the problem of high-dimensional parameter estimation. The key idea is to group measurements into clusters which are affected only by a subset of the parameter space. In contrast to static partitioning techniques, the method presented dynamically generates clusters at each step of the estimation. This substantially reduces the computation required, even for problems which cannot be partitioned in the traditional sense. Hence the resulting system is able to perform robust visual tracking of all six degrees of freedom in real-time. Small helicopters are extremely unstable and require high bandwidth active control. This in turn requires accurate estimates of both the pose and (importantly) velocity of the helicopter. Unfortunately, visual tracking gives noisy estimates of position; this, coupled with the non-Gaussian dynamic processes, means that conventional filtering techniques either yield unusably poor velocity estimates or respond insufficiently to large disturbances. In this thesis, a multiple hypothesis filter which employs a dynamic programming algorithm is used. This filter is responsive, while yielding stable and accurate estimates of pose and velocity. These estimates can then be used with a simple controller to reliably control the orientation of the helicopter. This controller is in turn ‘nested’ inside a controller which chooses the orientation required to reach or hold at a desired position. This combined controller is able to reliably fly the helicopter and results from a variety of test flights are presented.