Visual Measurement And Calculating Method Of UAV Position And Attitude Based On Deep Learning

With the development of drones,more and more drone products have appeared to meet the different needs of people.Due to the increasing demand for automation and intelligence of drone flight control,we need to accurately sense the position and attitude of the drone.The traditional methods usually have the disadvantages of large accumulated error and vulnerability to external interference.The measurement of the UAV by visual signals can not only achieve low-cost non-contact measurement of the target but also detect a large amount of effective information.In order to realize the UAV pose estimation,the method of measuring and calculating the position and attitude of the UAV using the visual sensor can effectively extract the key information in the image.This paper first establishes the coordinate system of visual measurement,and analyzes the projection imaging principle of the camera and the mathematical model contained in it.Camera imaging principle and binocular convergence model are the basis of pose estimation in this paper.To solve the parameters in the camera model,the internal and external parameters of the binocular camera are solved by using the calibration plate photo.Secondly,in order to extract key coordinate information from complex visual images,a keypoints prediction algorithm based on convolutional neural network for double convolutional networks is proposed.Key points are detected by the first convolutional neural network,based on predicted coordinates.The image segmentation and the key points of the second convolution network are optimized in three steps to finally output the information of the four rotor axis coordinates of the drone.In order to locate the matching pixels between the binocular cameras,the target centroid based on the Faster RCNN is used to locate and candidate region screening,ROI pooling size modification,and final loss function minimization.Thirdly,this paper uses the four feature points detected by the double convolution network,and inputs the PNP algorithm in the visual principle to calculate the space pose of the drone.The coordinates of the monocular image detection do not include the position information in the target space,and the binocular intersection model is established by the relationship between the binocular cameras,and the position information of the drone is finally output by using the binocular matching coordinates detected by the target detection network.Finally,the measurement method of the attitude and position of the drone is combined into a system,and the attitude detection system is tested experimentally.The experimental results show that the method is feasible and can meet the visual measurement of the drone.The innovation of this paper is reflected in the position and attitude measurement and calculation method using the visual key information detection method based on deep learning combined with the traditional visual imaging principle.This method can be used in the field of autonomous navigation,autonomous landing and automatic obstacle avoidance of drones.Get a wide range of applications.This research was funded by the National Natural Science Foundationof China(No.91646108,61473322).