Development of a visual-teach-and-repeat based navigation technique on quadrotor aerial vehicle

摘要

The objective of this thesis is to develop a vision-based navigation and control techniqueudfor quadrotor to operate in GPS-denied environments. The navigating techniqueudhas been developed while using Visual-Teach-and-Repeat (VT&R) method. Thisudmethod is qualitative where the position of the quadrotor is estimated based on a setudof reference images. These reference images are collected while taking the quadrotorudmanually along a desired route. Each image, collected in the database, represents oneudsegment of the desired route. The features are extracted from these images using audwell-known method, Speeded-Up Robust Features (SURF) [1].udWhen the quadrotor is navigated along the desired route (repeat mode), the quadrotorudperforms self-localization. Three methods of self-localization are presented. In methodudI, the SURF features observed on the current image are matched with the SURFudfeatures of the reference images to compute the probability value of each segment inudthe desired route. The segment that provides the best probability value is chosen asudthe current segment of the quadrotor. To improve the accuracy of localization, in theudmethod II, the condition of feature-size relation with spatial distance is imposed. Inudthe method III, the estimation of the current segment of the quadrotor is built onudBayes’s rule.udBased on the appearance-based error of feature coordinates, the system computesudqualitative motion control commands (desired yaw and height) for the next movement in order to control the quadrotor to follow the desired route. This computation isuddeveloped on Funnel Lane theory, which was originally proposed in [2], in order toud2D navigate ground vehicle following the desired route. The thesis extends it toud3D navigation for the quadrotor. Funnel Lane theory qualitatively defines possibleudpositions where the vehicle can fly straight by the constraints of features coordinatesudbetween the current image and the reference image. If the quadrotor locates outsideudthe funnel lane, it will be navigated back to the funnel lane.udA nonlinear geometric controller has been developed to convert the motion controludcommands, generated basing on VT&R technique, into control inputs necessary forudthe four rotors in the quadrotor. The design of proposed controller is simplified byudconcentrating on the errors of rotational matrix, instead of attempting to access theuderrors of each degree of freedom.udThe quadrotor for this thesis is chosen as the well-known AR.Drone model [3]. Theudwhole system is modeled and simulated in Gazebo simulator using Robot OperatingudSystem (ROS). Four image databases have been used for testing self-localization:udtwo databases around Engineering building of Memorial University of Newfoundland,udCOLD database and New College database. With proposed VT&R technique, theudquadrotor is able to independently follow a long route without GPS-information orudthe support from an external tracking system. The proposed system has a simpleudimplementation, inexpensive computation and high potential for exploring andudsearching-and-rescuing missions.