Vision based control of autonomous quadcopter for target tracking

Abstract

Conventional identification techniques for commercial quadcopters pose several shortcomings, such as limited system order, lack of statistical and non-parametric analysis, and not estimating the model’s linear operating range and quadcopter noise dynamics. This affects the prediction accuracy of quadcopter longitudinal and lateral motion dynamics that ultimately limits the quadcopter stabilization. To handle these challenges, in this thesis, statistically suitable plant and noise models are proposed for longitudinal and lateral motion dynamics of AR.Drone 2.0 quadcopter via the Box-Jenkins (BJ) model structure. Utilizing the flight data from the quadcopter, the models were estimated using Prediction Error Method (PEM) guided by statistical, non-parametric, and cross-validation analysis. The goodness of fit showed that the predicted model output matches the measured flight data by 94.72% in the one-step-ahead prediction test. When compared with first and second order models, the results revealed an improvement in prediction accuracy by 52.80%. In terms of image-based control of quadcopter translational dynamics, the rotation sensitivity of normalized spherical image features generates image feature errors and nonlinear coupling effects on the translational degrees of freedom. This causes unsuitable or unnecessary motions, thus affecting the positioning accuracy of the quadcopter. To overcome these limitations, this thesis proposes an image-based position control algorithm using rotation-invariant normalized spherical image features derived from a virtual spherical camera approach and optimally estimated using a Kalman filter method. For longitudinal and lateral translational motion control, the control system comprises an image-based outer-loop control law (developed using a proportional control action) cascaded with a velocity-based inner-loop control law (developed using a discrete-time proportional-integral-derivative (PID) control action). The control of vertical translational motion is based on image-based outer-loop control law. During the combined image-based positioning and hovering tasks, the proposed control algorithm regulates the image feature error in a maximum average time of approximately 25.29 s with a maximum average positioning accuracy of approximately 96.34%. For the combined image-based target tracking and hovering tasks, after the first disturbance of target object has vanished, the proposed control algorithm regulates the image feature error in a maximum average time of approximately 9.06 s. To further enhance the capability of the proposed control system, this thesis proposes an extremum seeking based automatic tuning system to determine the optimal vertical motion servoing gain that will optimize the response of filtered vertical motion image feature. This drastically improved the rise time and the time needed to reach the setpoint by approximately 57.82% and 59.22%, respectively, during image-based positioning tasks. The outcome from this research work had demonstrated adaptive image-based flight controllers which ultimately would be extensively useful for selfie drones, multiple target inspection tasks, and high-speed autonomous drone racing. Keywords: Quadcopter, PEM method, BJ model, image-based control, spherical image features, image moment features, Kalman filter, extremum seeking.