Research And Practice Of Pose Estimation In The Vision-Based Navigation For UAV

UAVs(Unmanned Aerial Vehicles)are becoming more and more intelligent.The visual system and computer vision are the main means to realize the intelligence of UAVs,especially in which the visual pose estimation is the fundamental technology for such intelligent systems as UAVs' environment perception,autonomous obstacle avoidance and navigation controlling.For the problem of pose estimation in UAVs' vision navigation system,taking the rotor UAV as experimental platform,the paper studies the key technologies of accurate pose estimation,vision-based autonomous landing and autonomous positioning in unknown environments.The main works and innovation points are as follows:1.To improve the efficiency of matching in sequence images,an algorithm of fast line detection and matching based on Hough one-dimensional transformation space and the constraints of point and line optical flows is proposed.The Hough one-dimensional space is used to reduce the complexity of computation and improve the accuracy of line detection.The fast tracking and matching of detected lines can be achieved with optical flow constraints.2.The problem of camera pose estimation with single image is studied.According to the different features used,it can be divided into two categories:PnP(Perspective-n-Point)and PnL(Perspective-n-Line).For the PnP problem,a pose estimation algorithm based on the minimization of object-space error and Lie group representation is proposed,which outperforms the traditional algorithms in efficiency,while the high positioning accuracy is kept.For the PnL problem,a pose estimation algorithm using coplanar line correspondences based on iteratively reweighted least squares is put forward,which can obtain the focal length and the pose of the camera simultaneously.In order to further improve the accuracy of PnL problem,an iterative algorithm based on the minimization of object-space coplanarity error of line and Lie group representation is proposed,which has few iterations,fast convergence and high accuracy.For the pose estimation in environments with sparse point features,an algorithm based on the fusion of point and line features is proposed to obtain high precision pose parameters.3.The problem of relative pose estimation of cameras between consecutive images is studied.Based on the epipolar geometric constraint,an iterative pose estimation algorithm using the improved essential matrix decomposition is proposed.The translation vector is determined firstly,the rotation matrix is obtained via the iterative method.In case of planar(or approximate planar)scene or short baseline,a direct pose estimation algorithm based on the minimization of homography mapping error and Lie group representation is proposed.With the known normalvector of the scene,the relative pose can be directly computed via feature correspondences,which avoids the processes of computing homography matrix and unique solution determination.Based on the geometric constraint among three images,an iterative pose estimation algorithm based on the trilinear algebraic error minimization is proposed,which can avoid the tedious computation of the trifocal tensor and improve the pose estimation robustness by using the three-view geometric constraint.Experiments show that the proposed algorithm can effectively improve the accuracy and efficiency of the relative pose estimation.4.A two-stage vision-based autonomous landing scheme is designed,which solves the problem of UAV pose estimation accuracy at different altitude.The process of landing is divided into two stages based on the altitude of UAVs.When the altitude of UAV is high and the target is small on image,the pose estimation method based on the cooperative target and epipolar geometry between two views is utilized,which can directly get the absolute scale of translation without the help of additional sensors.When the UAV is close to the cooperative target,the pose estimation method based on single image can obtain the relative poses between UAV and cooperative target.In order to further improve the efficiency,an algorithm based on the combination of line features and extended Kalman filter is proposed.Experimental results show that the proposed two-stage landing scheme can obtain the UAV's poses with high accuracy.5.To realize the UAVs' autonomous positioning in unknown environments,a pose estimation algorithm based on three-view is proposed.The method takes three consecutive images as a processing unit.Firstly,pyramid optical flow tracking and forward and backward error scheme are utilized to realize the fast features detection and matching of high resolution images.Secondly,the overlapping images between processing units are used to achieve the scale consistency,and further the global scale consistency of relative pose can be realized.A scale correcting strategy based on the known height of camera is designed to rectify the scale drift.The proposed algorithm avoids the reconstruction of the scene structure in traditional methods,which can reduce the complexity of the algorithm,save the limited storage resources,and thus is suitable for the UAV platform.Experimental results show that the proposed method can process the UAV high-resolution image sequences in real-time and maintain the accuracy of UAV's pose estimation over a long time and large flight range.