Research On Vision-based Pose Estimation And Object Tracking For Miniature Unmanned Helicopter

Unmanned Aerial Vehicles (UAVs), which are being widely used in both military and civil applications, have been one of the most hot research topics global wide in recent years. The implementation of autonomous navigation is the prerequisite for UAV to be applied in reality. In comparison with traditional navigation that relies on inertial sensors and global position system (GPS), vision based navigation always acquires environmental information in real time, and can manage tasks related with environment such as relative pose estimation, obstacle avoidance maneuver, surveillance and tracking, which makes it receive more and more attention. In this dissertation, a miniature unmanned helicopter (MUH) is selected as the main experimental platform, and some researches have been done to solve problems in vision-based navigation, including relative pose estimation and ground target tracking. The main work of this dissertation is listed below:In chapter1, research background and significance of this dissertation are introduced. Firstly, domestic and foreign research on vision-based navigation of UAV is reviewed. Then, the vision based navigation methods are classified according to the theories they use, and both the hardware platform and software system for ground target tracking are analyzed. Finally, the research contents and organization of this dissertation are clarified.In chapter2, a brief introduction to the related basic theories is made. Firstly, the usually used reference coordination definitions and attitude description methods are described. Then, the relative knowledge of projective geometry, camera model and color image is introduced. Finally, the simultaneous localization and mapping algorithm based on extended Kalman filter (EKF-SLAM) is discussed.In chapter3, in order to estimate the attitude of MUH flying in urban environment, a method to extract the vanishing point of vertical edge lines belong to buildings in aerial images captured by downwards looking camera is proposed. This method firstly extracts edge lines using Hough transform, and then extracts the vanishing point and vertical edge lines simultaneously using inverse Hough transform. The attitude of MUH is directly calculated from vanishing point’s coordinate according to projective geometry. Experimental results show that this method can extract vanishing point reliably, and the estimated attitude is accurate. In chapter4, in order to estimate the attitude of high altitude flying MUH, a method to extract horizon line in aerial images captured by forwards looking camera is proposed. This method firstly extracts the longest lines as the candidates of horizon using Hough transform, and then determines the horizon by taking region information of the dark channel priori image into consideration. Attitude of MUH is estimated using an EKF that combines horizon observation model and rigid body rotation model. The predicted horizon’s parameters and innovation are used for judging whether the extracted horizon is true. Correction will be carried out when the extracted horizon line is erroneous. Experimental results show that this method can extract horizon robustly.In chapter5, in order to estimate the attitude of MUH hovering and rotating, modifications are made to a MonoSLAM-based attitude estimation method called "visual compass". Since MUH has large angular accelerations, the visual compass method can work only if the a priori system noise covariance in motion model is set a large value. But this results in high computational cost for matching and high rate of erroneous matches. A multi resolution landmark selection strategy is adopted for initializing new landmark. Correspondingly, an active search and match layer by layer algorithm is proposed for fast landmark matching. The modified visual compass method is proved superior in less computational cost and lower rate of erroneous matches by experimental results.In chapter6, in order to estimate the pose of MUH flying close to ground, modifications are made to the pose estimation method that based on MonoSLAM. Instead of parameterising several simultaneously initialized landmarks individually, landmark clutter is used for less camera parameter redundancy and reinforced camera pose constraints. The initial depth of landmark is set a larger value for deducing the ratio of changing depth from positive to negative after EKF updating. A2point random sample consensus (RANSAC) data association algorithm is proposed to cope with situation when angular velocities around multi-axis of camera change rapidly. The accuracy of modified MonoSLAM-based pose estimation method fulfills the requirement for autonomous flight in field-experiment.In chapter7, in order to track the ground object from a MUH, a hardware platform for surveillance and tracking is constructed. Two gimbals are designed and manufactured. The first one with two-axis two-frame is traditional, and the second one with two-axis three-frame can avoids system singularity encountered in two-axis two-frame gimbals. Besides, a visual tracking algorithm based on mean shift and elliptical blob detection in likelihood map is proposed. This algorithm uses object/mixed regions’area weighted ratio of histograms as likelihood to suppress the interference from background. In likelihood map, the object appears to be an elliptical blob. The position of blob is estimated using mean shift, while the orientation, semi-major axis and semi-minor axis of blob are determined by calculating the difference of elliptical Gaussian. Experimental results show that this algorithm can track object with both rotation and scale changes.In chapter8, the current research work of this dissertation is summarized and some recommendations for further research are presented.