Research On Key Technologies Of Visual Aids-based Landing In Unknown Area For Unmanned Aerial Vehicle

Unmanned Aerial Vehicle(UAV)has to face the autonomous landing problem of complex terrain and no ground assisted navigation equipment,when it is used for rescuing and searching.It plays an important role in improving the security of UAV in the unknown area by using computer vision to assist UAV searching for the safe landing area and guide it to land autonomously.Some key technologies in visual aids for an unknown area of the UAV landing are studied in this dissertation,the main research contents and contributions are as follows:A measurement of the UAV’s height obtained by a single sensor is usually not accurate and easily disturbed.A method of adaptive S-filter for height information fusion is developed to solve the problem.Firstly,an adaptive S-filter is designed to deal with the noise existed in the original data from the sensor;Secondly,the height measurement model of each sensor is established,and the centralized Kalman filter is used to smooth the data.So the accurate height estimation of UAV can be gotten.Finally,the proposed method is compared with multi-sensor fusion and the fusion method after wavelet filter.The simulation results show that the proposed method can eliminate the noise effectively and obtain higher estimation accuracy.The acquired aerial image in unknown area is susceptible to illumination and scale changes which easily cause the low segmentation accuracy for the terrain.To solve the problem,an adaptive segmentation and searching algorithm for the landing area with multi-feature fusion and assisted by altitude is proposed.Firstly,the needed minimum pixel is calculated for UAV landing according to the current height of UAV,the size of the landing area and the ground resolution of the image.Secondly,the Mean Shift algorithm is employed for image coarse segmentation,and the bandwidth of the kernel function in Mean Shift is calculated according to the minimum pixel gotten in previous step in combination with the threshold in maximum between-cluster variance.Thirdly,the edge of the coarse segmentation image is drawn by using the Canny operator and the landing areas which meet landing requirements are found through the propose landing area template searching strategy.Finally,aerial images of different scenarios and scales selected from Google earth and the aerial images prepared by UAV are used for segmentation experiment.Experimental results demonstrate that the proposed algorithm can meet the mission requirements for accurate segmentation and it is robust to illumination and scale change.The acquired aerial image is susceptible to the illumination and noises which easily cause poor classification performance and weak robustness.To address this issue,a low rank and sparse representation terrain multi-stage classification method oriented to aerial image based on feature dictionary is proposed.Firstly,the sample images are transformed to HSV space from RGB space,and color moments and Gabor texture feature are extracted from training data and stacked as column vectors of a dictionary.Secondly,an Augmented Lagrange Multipliers(ALM)algorithm is used to decompose the feature dictionary into a low rank matrix and a sparse matrix.And a multi-stage classifier is constructed on the basis of the low rank matrix.Finally,the effectiveness of the algorithm is verified based on aerial image database prepared by our own.Experimental results demonstrate that the proposed algorithm has high classification accuracy and strong robustness.To solve the problem that the large error of depth information reconstruction exists in sparse matching and high false match rate based on dense matching is high in the smooth region,a dense point feature generation algorithm based on monocular sequence images for depth measurement of unknown area.Firstly,sub pixel Harris corner and scale invariant feature transform(SIFT)feature points are extracted and matched respectively in two frames which are selected from sequence images.Secondly,the two type feature points are fused under the conditions of Euclidean distance between them.So quasi dense feature points can be obtained.Thirdly,quasi dense feature points are Delaunay triangulated and dense feature points generating strategy is developed according to the variance of the three vertex pixel deviation in each triangulation triangle.Depth information of the whole unknown area is calculated according to the proposed depth calculation equation.Finally,the proposed algorithm is tested on Vega Prime(VP)and the experimental platform which is composed of UAV and the cooperative target.Experimental results demonstrate that the proposed algorithm has high depth estimation precision.In order to overcome the problems that randomly selecting feature points for relative position and angle estimation leads to low precision estimation and poor stability,a selection algorithm of random feature points based on vector constraints is proposed.Firstly,the position and pose estimation equations on the basis of the principle of the pinhole imaging principle is converted to a standard linear equations.Through the analysis,condition number of the linear equations is mainly decided by the geographic coordinates of feature points.So the feature points selection directly influences the precision of relative position and angle estimation.Based on this,the vector angle average degree,the mean of vector modulus and the maximum value of vector modulus are introduced for evaluating the influence of thefeature points on the accuracy of position and angle estimation.Secondly,selection strategy of random feature points based on vector constraints is developed.Finally,on the basis of orthogonal iterative algorithm,the propose algorithm is tested on simulation and physical experiment platform respectively.The experimental results show that the proposed algorithm has higher accuracy and stronger robustness compared to the method of randomly selecting feature points.