Vision-based target tracking architecture for an unmanned aerial vehicle

Abstract

This thesis presents target tracking architecture for an unmanned aerial vehicle using onboard pan-tilt camera system. The target tracking architecture is divided into four parts: image stabilization, color-based object tracking, target geo-location, visual servoing. Video acquired using an onboard camera on a small unmanned aerial vehicle is commonly plagued with distractive jittery motions and disoriented rotations. These problems make it very difficult for human viewer to identify and select objects of interest. In order to stabilize UAV video, a real-time feature based image stabilization method is proposed. For tracking target selected by user color-based object tracking method is used in this thesis. The target histogram model is weighted by monotonically decreasing kernel profile to reduce influence of background prior to histogram back-projection of probability density image. Using CamShift algorithm on probability image, the target is tracked and new size and orientation of target is updated in each image sequence. Using the location of target in an image, position and attitude of a vehicle, and pan and tilt angle, the object on ground is localized in world coordinates with flat earth model. The guidance law of vehicle is proposed based on position-based visual servoing. The camera gimbal system is manipulated based on image-based visual servoing and provides sustained object acquisition and rapid tracking capability. Through real-time image processing using video acquired on RUAV platform and non-linear helicopter simulation, these tracking architecture was shown to provide a powerful vision-based target tracking for an unmanned aerial vehicle with onboard pan-tilt camera system.